├── .gitignore ├── .idea ├── CrosswalkDetectionNet.iml ├── codeStyles │ ├── Project.xml │ └── codeStyleConfig.xml ├── deployment.xml ├── dictionaries │ └── tanmenglu.xml ├── inspectionProfiles │ └── Project_Default.xml ├── misc.xml ├── modules.xml ├── remote-mappings.xml ├── vagrant.xml ├── vcs.xml └── workspace.xml ├── README.md ├── __init__.py ├── data ├── Detection results compare to YOLOv5.jpg ├── Detection results compare to YOLOv5_0.jpg ├── Detection results compare to YOLOv5_1.jpg ├── SHNAVI.jpg ├── cl_setting.yaml ├── crosswalk detection in complex realistic scenarios.jpg ├── crosswalk.yaml ├── graphical_abstract.jpg ├── rect.jpg └── set_ai.png ├── detect.py ├── docs ├── DATASETS.md ├── INSTALL.md ├── MODEL_ZOO.md ├── Press_Material.md └── README_zh_cn.md ├── evaluate.py ├── example ├── fog_image │ ├── fogged_br0.1_th0.05.jpg │ ├── fogged_br0.3_th0.05.jpg │ ├── fogged_br0.5_th0.05.jpg │ ├── fogged_br0.7_th0.01.jpg │ ├── fogged_br0.7_th0.03.jpg │ ├── fogged_br0.7_th0.05.jpg │ ├── fogged_br0.7_th0.07.jpg │ ├── fogged_br0.7_th0.09.jpg │ ├── fogged_br0.9_th0.05.jpg │ └── raw.jpg ├── images │ ├── complex_scenarios │ │ ├── clearly_visible_0760_00001.jpg │ │ ├── damaged_0036_00846.jpg │ │ ├── dazzling_0036_01707.jpg │ │ ├── distorted_0760_01846.jpg │ │ ├── occluded_by_wiper_0760_02902.jpg │ │ ├── partialy_lost_0036_01558.jpg │ │ ├── reflection_after_rain_0760_02711.jpg │ │ ├── truncated_in_steering_0760_03208.jpg │ │ ├── view_blocked_0036_05420.jpg │ │ └── visible_at_night_0036_01127.jpg │ ├── exmp_imgs │ │ ├── filename_00030.jpg │ │ ├── filename_00031.jpg │ │ ├── filename_00032.jpg │ │ ├── filename_00033.jpg │ │ ├── filename_00034.jpg │ │ ├── filename_00035.jpg │ │ ├── filename_00036.jpg │ │ ├── filename_00037.jpg │ │ ├── filename_00038.jpg │ │ ├── filename_00039.jpg │ │ ├── filename_00040.jpg │ │ ├── filename_00041.jpg │ │ ├── filename_00042.jpg │ │ ├── filename_00043.jpg │ │ ├── filename_00044.jpg │ │ ├── filename_00045.jpg │ │ ├── filename_00046.jpg │ │ ├── filename_00047.jpg │ │ ├── filename_00048.jpg │ │ ├── filename_00049.jpg │ │ ├── filename_00050.jpg │ │ ├── filename_00051.jpg │ │ ├── filename_00052.jpg │ │ ├── filename_00053.jpg │ │ ├── filename_00054.jpg │ │ ├── filename_00055.jpg │ │ ├── filename_00056.jpg │ │ ├── filename_00057.jpg │ │ ├── filename_00058.jpg │ │ └── filename_00059.jpg │ ├── filename_00030.jpg │ ├── filename_00031.jpg │ ├── filename_00032.jpg │ ├── filename_00033.jpg │ ├── filename_00034.jpg │ ├── filename_00035.jpg │ ├── filename_00036.jpg │ ├── filename_00037.jpg │ ├── filename_00038.jpg │ ├── filename_00039.jpg │ ├── filename_00040.jpg │ ├── filename_00041.jpg │ ├── filename_00042.jpg │ ├── filename_00043.jpg │ ├── filename_00044.jpg │ ├── filename_00045.jpg │ ├── filename_00046.jpg │ ├── filename_00047.jpg │ ├── filename_00048.jpg │ ├── filename_00049.jpg │ ├── filename_00050.jpg │ ├── filename_00051.jpg │ ├── filename_00052.jpg │ ├── filename_00053.jpg │ ├── filename_00054.jpg │ ├── filename_00055.jpg │ ├── filename_00056.jpg │ ├── filename_00057.jpg │ ├── filename_00058.jpg │ ├── filename_00059.jpg │ └── for_overall_architecture │ │ ├── filename1120.jpg │ │ ├── filename1121.jpg │ │ ├── filename1122.jpg │ │ ├── filename1123.jpg │ │ ├── filename1124.jpg │ │ ├── filename1125.jpg │ │ ├── filename1126.jpg │ │ ├── filename1127.jpg │ │ ├── filename1128.jpg │ │ ├── filename1129.jpg │ │ ├── filename1130.jpg │ │ └── filename1131.jpg ├── out_for_overall │ ├── detect.log │ ├── filename1120.jpg │ ├── filename1120.txt │ ├── filename1121.jpg │ ├── filename1121.txt │ ├── filename1122.jpg │ ├── filename1122.txt │ ├── filename1123.jpg │ ├── filename1123.txt │ ├── filename1124.jpg │ ├── filename1124.txt │ ├── filename1125.jpg │ ├── filename1125.txt │ ├── filename1126.jpg │ ├── filename1126.txt │ ├── filename1127.jpg │ ├── filename1127.txt │ ├── filename1128.jpg │ ├── filename1128.txt │ ├── filename1129.jpg │ ├── filename1129.txt │ ├── filename1130.jpg │ ├── filename1130.txt │ ├── filename1131.jpg │ └── filename1131.txt └── output │ ├── detect.log │ ├── filename_00030.jpg │ ├── filename_00031.jpg │ ├── filename_00032.jpg │ ├── filename_00033.jpg │ ├── filename_00034.jpg │ ├── filename_00035.jpg │ ├── filename_00036.jpg │ ├── filename_00036.txt │ ├── filename_00037.jpg │ ├── filename_00037.txt │ ├── filename_00038.jpg │ ├── filename_00038.txt │ ├── filename_00039.jpg │ ├── filename_00039.txt │ ├── filename_00040.jpg │ ├── filename_00040.txt │ ├── filename_00041.jpg │ ├── filename_00041.txt │ ├── filename_00042.jpg │ ├── filename_00042.txt │ ├── filename_00043.jpg │ ├── filename_00043.txt │ ├── filename_00044.jpg │ ├── filename_00045.jpg │ ├── filename_00046.jpg │ ├── filename_00046.txt │ ├── filename_00047.jpg │ ├── filename_00047.txt │ ├── filename_00048.jpg │ ├── filename_00048.txt │ ├── filename_00049.jpg │ ├── filename_00049.txt │ ├── filename_00050.jpg │ ├── filename_00050.txt │ ├── filename_00051.jpg │ ├── filename_00052.jpg │ ├── filename_00053.jpg │ ├── filename_00054.jpg │ ├── filename_00055.jpg │ ├── filename_00056.jpg │ ├── filename_00057.jpg │ ├── filename_00058.jpg │ └── filename_00059.jpg ├── fog_augment.py ├── for_paper ├── CDNet_plot.py ├── data │ └── results.yaml ├── entropy_analyse.py └── overall_data ├── models ├── common.py ├── commonSE.py ├── experimental.py ├── yolo.py ├── yoloSE.py └── yolov5m.yaml ├── repos └── mmdetection │ ├── configs │ ├── _base_ │ │ └── datasets │ │ │ ├── zzd_coco_detection.py │ │ │ ├── zzd_coco_instance.py │ │ │ ├── zzd_crosswalk_detection.py │ │ │ └── zzd_crosswalk_instance.py │ ├── centernet │ │ └── zzd_centernet_resnet18_dcnv2_140e_crosswalk.py │ ├── deformable_detr │ │ └── zzd_deformable_detr_r50_16x2_50e_crosswalk.py │ ├── faster_rcnn │ │ ├── zzd_faster_rcnn_r50_fpn_1x_coco.py │ │ └── zzd_faster_rcnn_r50_fpn_1x_crosswalk.py │ ├── htc │ │ ├── zzd_htc_r50_fpn_1x_crosswalk.py │ │ └── zzd_htc_without_semantic_r50_fpn_1x_crosswalk.py │ ├── mask_rcnn │ │ ├── zzd_mask_rcnn_r50_fpn_1x_coco.py │ │ └── zzd_mask_rcnn_r50_fpn_1x_crosswalk.py │ ├── scnet │ │ └── zzd_scnet_r50_fpn_1x_crosswalk.py │ ├── swin │ │ └── zzd_mask_rcnn_swin-t-p4-w7_fpn_1x_crosswalk.py │ └── yolox │ │ ├── zzd_yolox_m_8x8_300e_crosswalk.py │ │ └── zzd_yolox_s_8x8_300e_crosswalk.py │ ├── demo │ └── zzd_inference_speed_test.py │ ├── zzd_inference.py │ └── zzd_train.py ├── requirements.txt ├── scripts ├── __init__.py ├── check_coco.py ├── check_dataset.py ├── get_confusing_matrix.py ├── post.bak.py ├── post.py ├── pt2wts.py ├── synthetic_fog.py ├── test.py └── yolo2coco.py ├── train.py └── utils ├── __init__.py ├── datasets.py ├── general.py ├── google_utils.py └── torch_utils.py /.gitignore: -------------------------------------------------------------------------------- 1 | /for_paper/ 2 | /repos 3 | /weights -------------------------------------------------------------------------------- /.idea/CrosswalkDetectionNet.iml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | -------------------------------------------------------------------------------- /.idea/codeStyles/Project.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 7 | 8 | 9 | -------------------------------------------------------------------------------- /.idea/codeStyles/codeStyleConfig.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 5 | -------------------------------------------------------------------------------- /.idea/deployment.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 15 | -------------------------------------------------------------------------------- /.idea/dictionaries/tanmenglu.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | -------------------------------------------------------------------------------- /.idea/inspectionProfiles/Project_Default.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 26 | -------------------------------------------------------------------------------- /.idea/misc.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 6 | 7 | -------------------------------------------------------------------------------- /.idea/modules.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | -------------------------------------------------------------------------------- /.idea/remote-mappings.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | -------------------------------------------------------------------------------- /.idea/vagrant.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 7 | -------------------------------------------------------------------------------- /.idea/vcs.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | [![Stars](https://img.shields.io/github/stars/zhangzhengde0225/CDNet)]( 2 | https://github.com/zhangzhengde0225/CDNet) 3 | [![Open issue](https://img.shields.io/github/issues/zhangzhengde0225/CDNet)]( 4 | https://github.com/zhangzhengde0225/CDNet/issues) 5 | [![Datasets](https://img.shields.io/static/v1?label=Download&message=datasets&color=green)]( 6 | https://github.com/zhangzhengde0225/CDNet/blob/master/docs/DATASETS.md) 7 | [![Datasets](https://img.shields.io/static/v1?label=Download&message=source_code&color=orange)]( 8 | https://github.com/zhangzhengde0225/CDNet/archive/refs/heads/master.zip) 9 | 10 | #### English | [简体中文](https://github.com/zhangzhengde0225/CDNet/blob/master/docs/README_zh_cn.md) 11 | 12 | Please **star this project** in the upper right corner and **cite this paper** blow 13 | if this project helps you. 14 | 15 | # CDNet 16 | 17 | This repository is the codes, datasets and tutorials for the paper 18 | "CDNet: a real-time and robust crosswalk detection network on Jetson nano based on YOLOv5". 19 | 20 | CDNet (Crosswalk Detection Network) is a specific implementation of crosswalk (zebra crossing) detection and vehicle crossing behavior analysis under the vision of vehicle-mounted camera. 21 | 22 | ![GA](https://zhangzhengde0225.github.io/images/CDNet_GA.jpg) 23 | 24 | Fig.1 Graphical abstract. 25 | 26 | # Highlights 27 | + A crosswalk detection and vehicle crossing behavior detection network is realized. 28 | + The accuracy and speed exceed YOLOv5 in the specific task. 29 | + High robustness in real complex scenarios such as in cloudy, sunny, rainy and at night is achieved. 30 | + Real-time detection (33.1 FPS) is implemented on Jetson nano edge-computing device. 31 | +The datasets, tutorials and source codes are available on GitHub. 32 | 33 | # Contribution 34 | 35 | + SENet (Squeeze-and-Excitation Network), F1 score up, speed slightly down 36 | + NST (Negative Samples Training), F1 score up, speed invariant 37 | + ROI (Region Of Interest), F1 score down, speed up 38 | + SSVM (Slide receptive field Short-term Vectors Memory), transfer crosswalk detection task into vehicle crossing behavior task, 39 | F1 score up, speed invariant 40 | + SFA (Synthetic Fog Augment), dataset augment, adapt to foggy weather, F1 score up, speed invariant 41 | 42 | # Installation 43 | Get CDNet code and configure the environment, please check out [docs/INSTALL.md](https://github.com/zhangzhengde0225/CDNet/blob/master/docs/INSTALL.md) 44 | 45 | # Model Zoo 46 | Please check out [docs/MODEL_ZOO.md](https://github.com/zhangzhengde0225/CDNet/blob/master/docs/MODEL_ZOO.md) 47 | 48 | # Datasets 49 | Download trainsets and testsets, please check out [docs/DATASETS.md](https://github.com/zhangzhengde0225/CDNet/blob/master/docs/DATASETS.md) 50 | 51 | # Quick Start 52 | ## Train 53 | 54 | Once you get the CDNet code, configure the environment and download the dataset, juse type: 55 | ``` 56 | python train.py --trainset_path 57 | (such as: /home/xxx/datasets/train_data_yolov5_format) 58 | ``` 59 | The training results and weights will be saved in runs/expxx/ directory. 60 | 61 | The main optional arguments: 62 | ``` 63 | --device "0, 1" # cpu or gpu id, "0, 1" means use two gpu to train. 64 | --img-size 640 65 | --batch-size 32 66 | --epochs 100 67 | --not-use-SE # use original YOLOv5 which not SE-module embedded if there is the flag 68 | ``` 69 | 70 | ## Inference 71 | Detect the crosswalk image by image and analyze the vehicle crossing behavior. 72 | ``` 73 | python detect.py 74 | ``` 75 | 76 | The main optional arguments: 77 | ``` 78 | --source example/images # images dir 79 | --output example/output # output dir 80 | --img-size 640 # inference model size 81 | --device "0" # use cpu or gpu(gpu id) 82 | --plot-classes ["crosswalk"] # plot classes 83 | --field-size 5 # the Slide receptive field size of SSVM 84 | --not-use-ROI # not use roi for accelerate inference speed if there is the flag 85 | --not-use-SSVM # not use ssvm method for analyse vehicle crossing behavior if there is the flag 86 | ``` 87 | 88 | For more details, please refer to [docs/INSTALL.md](https://github.com/zhangzhengde0225/CDNet/blob/master/docs/INSTALL.md) and [docs/DATASETS.md](https://github.com/zhangzhengde0225/CDNet/blob/master/docs/DATASETS.md). 89 | 90 | ## Fogging Augment 91 | If you want to augment datasets by synthetic fog algorithm, just run: 92 | ``` 93 | python fog_augment.py 94 | ``` 95 | For more details, please view the source code in fog_augment.py and /scripts/synthetic_fog.py 96 | 97 | # Results 98 | 99 | ![Results](https://zhangzhengde0225.github.io/images/CDNet_Results.jpg) 100 | 101 | Fig.2 Performance compared to YOLOv5. 102 | 103 | **CDNet improves the score for 5.13 points and speed for 10.7 FPS on Jetson nano for detection size of 640 compared to YOLOv5.** 104 | 105 | **For detection size of 288, the improvements are 13.38 points and 13.1 FPS.** 106 | 107 | 108 | # Contributors 109 | CDNet is authored by Zhengde Zhang, Menglu Tan, Zhicai Lan, Haichun Liu, Ling Pei and Wenxian Yu. 110 | 111 | Currently, it is maintained by Zhengde Zhang (drivener@163.com). 112 | 113 | Please feel free to contact us if you have any question. 114 | 115 | The Academic homepage of Zhengde Zhang: [zhangzhengde0225.github.io](https://zhangzhengde0225.github.io). 116 | 117 | # Acknowledgement 118 | 119 | This work was supported by the National Natural 120 | Science Foundation of China [Grant Numbers: 61873163]. 121 | 122 | We acknowledge the Center for High Performance Computing at 123 | Shanghai Jiao Tong University for providing computing resources. 124 | 125 | We are very grateful to the 126 | [yolov5](https://github.com/ultralytics/yolov5) project 127 | for the benchmark detection algorithm. 128 | 129 | We are very grateful to the 130 | [tensorrtx](https://github.com/wang-xinyu/tensorrtx) project 131 | for the deployment techniques to the Jetson nano. 132 | 133 | # Links 134 | Detect Video Samples：[https://www.bilibili.com/video/BV1qf4y1B7BA](https://www.bilibili.com/video/BV1qf4y1B7BA) 135 | 136 | Read Full Text of This Paper：[https://rdcu.be/cHuc8](https://rdcu.be/cHuc8) 137 | 138 | Download Full Text of this Paper：[https://doi.org/10.1007/s00521-022-07007-9](https://doi.org/10.1007/s00521-022-07007-9) 139 | 140 | Project Introduction on CSDN：[http://t.csdn.cn/Cf7c7](http://t.csdn.cn/Cf7c7) 141 | 142 | If it is helps you, 143 | please star this project in the upper right corner and cite this paper blow. 144 | 145 | # Citation 146 | ``` 147 | @article{CDNet, 148 | author={Zheng-De Zhang, Meng-Lu Tan, Zhi-Cai Lan, Hai-Chun Liu, Ling Pei and Wen-Xian Yu}, 149 | title={CDNet: a real-time and robust crosswalk detection network on Jetson nano based on YOLOv5}, 150 | Journal={Neural Computing and Applications}, 151 | Year={2022}, 152 | DOI={10.1007/s00521-022-07007-9}, 153 | } 154 | ``` 155 | 156 | # License 157 | CDNet is freely available for free non-commercial use, and may be redistributed under these conditions. For commercial queries, please drop an e-mail at drivener@163.com. We will send the detail agreement to you. 158 | 159 | 160 | 161 | 162 | 163 | -------------------------------------------------------------------------------- /__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/__init__.py -------------------------------------------------------------------------------- /data/Detection results compare to YOLOv5.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/data/Detection results compare to YOLOv5.jpg -------------------------------------------------------------------------------- /data/Detection results compare to YOLOv5_0.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/data/Detection results compare to YOLOv5_0.jpg -------------------------------------------------------------------------------- /data/Detection results compare to YOLOv5_1.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/data/Detection results compare to YOLOv5_1.jpg -------------------------------------------------------------------------------- /data/SHNAVI.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/data/SHNAVI.jpg -------------------------------------------------------------------------------- /data/cl_setting.yaml: -------------------------------------------------------------------------------- 1 | # control line setting for vehicle crossing crosswalk behavior detection 2 | 3 | # control_line in pixels for different videos 4 | general: [400, 680] # two horizontal lines parallel to the y-axis in pixels coordinates. [top, button] 5 | '01_4885': [380, 718] 6 | '02_0036': [400, 680] 7 | '03_0760': [380, 718] 8 | '04_1421': [400, 680] 9 | '05_0823': [225, 590] 10 | '06_5271': [365, 635] 11 | '13_3519': [300, 718] 12 | '14_9247': [280, 718] 13 | '15_9440': [270, 718] 14 | 15 | 16 | # red line 17 | red_line: 0.5 # proportion of diff. between red_line and top_line to diff. between button_line and top_line 18 | 19 | -------------------------------------------------------------------------------- /data/crosswalk detection in complex realistic scenarios.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/data/crosswalk detection in complex realistic scenarios.jpg -------------------------------------------------------------------------------- /data/crosswalk.yaml: -------------------------------------------------------------------------------- 1 | # PASCAL VOC dataset http://host.robots.ox.ac.uk/pascal/VOC/ 2 | # Download command: bash ./data/get_voc.sh 3 | # Train command: python train.py --data voc.yaml 4 | # Default dataset location is next to /yolov5: 5 | # /parent_folder 6 | # /VOC 7 | # /yolov5 8 | 9 | 10 | # train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/] 11 | # train: /home/zzd/datasets/crosswalk/train_data_v5_format/images/train 12 | # val: /home/zzd/datasets/crosswalk/train_data_v5_format/images/test 13 | 14 | # number of classes 15 | nc: 2 16 | 17 | # class names 18 | names: ['crosswalk', 'guide_arrows'] 19 | -------------------------------------------------------------------------------- /data/graphical_abstract.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/data/graphical_abstract.jpg -------------------------------------------------------------------------------- /data/rect.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/data/rect.jpg -------------------------------------------------------------------------------- /data/set_ai.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/data/set_ai.png -------------------------------------------------------------------------------- /detect.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import os 3 | import platform 4 | import shutil 5 | import time 6 | from pathlib import Path 7 | import numpy as np 8 | import yaml 9 | 10 | import cv2 11 | import torch 12 | import torch.backends.cudnn as cudnn 13 | from numpy import random 14 | from tqdm import tqdm 15 | 16 | from models.experimental import attempt_load 17 | from utils.datasets import LoadStreams, LoadImages 18 | from utils.general import check_img_size, non_max_suppression, apply_classifier, scale_coords, xyxy2xywh, plot_one_box 19 | from utils.torch_utils import select_device, load_classifier, time_synchronized 20 | from scripts import post 21 | 22 | 23 | def detect(opt, dp, save_img=False): 24 | out, source, weights, view_img, save_txt, imgsz = \ 25 | opt.output, opt.source, opt.weights, opt.view_img, opt.save_txt, opt.img_size 26 | webcam = source == '0' or source.startswith('rtsp') or source.startswith('http') or source.endswith('.txt') 27 | 28 | # Initialize 29 | device = select_device(opt.device) 30 | if os.path.exists(out): 31 | shutil.rmtree(out) # delete output folder 32 | os.makedirs(out) # make new output folder 33 | half = device.type != 'cpu' # half precision only supported on CUDA 34 | 35 | # Load model 36 | model = attempt_load(weights, map_location=device) # load FP32 model 37 | 38 | imgsz = check_img_size(imgsz, s=model.stride.max()) # check img_size 如果不是32的倍数，就向上取整调整至32的倍数并答应warning 39 | 40 | if half: 41 | model.half() # to FP16 42 | 43 | # Second-stage classifier 44 | classify = False 45 | if classify: 46 | modelc = load_classifier(name='resnet101', n=2) # initialize 47 | modelc.load_state_dict(torch.load('weights/resnet101.pt', map_location=device)['model']) # load weights 48 | modelc.to(device).eval() 49 | 50 | # Set Dataloader 51 | vid_path, vid_writer = None, None 52 | if opt.use_roi: 53 | # print(dp.cl) 54 | # print(dp.cl[0], dp.cl[1]) 55 | # cl = opt.control_line 56 | cl = dp.cl 57 | roi_in_pixels = np.array([0, cl[0], 1280, cl[1]]) # two points coor, x1, y1, x2, y2 58 | else: 59 | roi_in_pixels = None 60 | 61 | if webcam: 62 | view_img = True 63 | cudnn.benchmark = True # set True to speed up constant image size inference 64 | dataset = LoadStreams(source, img_size=imgsz) 65 | else: 66 | save_img = True 67 | dataset = LoadImages(source, img_size=imgsz, roi=roi_in_pixels) 68 | 69 | # Get names and colors 70 | names = model.module.names if hasattr(model, 'module') else model.names # 解决GPU保存的模型多了module属性的问题 71 | colors = [[random.randint(0, 255) for _ in range(3)] for _ in range(len(names))] # 随机颜色，对应names，names是class 72 | 73 | # fix issue: when single cls, names = ['item'] rather than names = ['crosswalk'] 74 | if 'item' in names: 75 | names = ['crosswalk'] 76 | 77 | # prune 78 | # torch_utils.prune(model, 0.7) 79 | model.eval() 80 | 81 | # Run inference 82 | t0 = time.time() 83 | img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img 84 | _ = model(img.half() if half else img) if device.type != 'cpu' else None # run once 空跑一次，释放！！牛逼 85 | 86 | detected_img_id = 0 87 | time_list = [None] * len(dataset) 88 | bar = tqdm(dataset) 89 | for iii, (path, img, im0s, vid_cap, recover) in enumerate(bar): 90 | # print(img.shape, im0s.shape, vid_cap) 91 | # exit() 92 | 93 | # img.shape [3, 384, 640] im0s.shape [720, 1280, 3] None 94 | img = torch.from_numpy(img).to(device) 95 | img = img.half() if half else img.float() # uint8 to fp16/32 96 | img /= 255.0 # 0 - 255 to 0.0 - 1.0 97 | if img.ndimension() == 3: 98 | img = img.unsqueeze(0) # 从[3, h, w]转换为[batch_size, 3, h, w]的形式 99 | 100 | # Inference 101 | t1 = time_synchronized() 102 | # print('aug', opt.augment) # False 103 | pred = model(img, augment=opt.augment)[0] 104 | # print(pred.shape) [1, 15120, 25] 105 | # Apply NMS 106 | pred = non_max_suppression(pred, opt.conf_thres, opt.iou_thres, classes=opt.classes, agnostic=opt.agnostic_nms) 107 | t2 = time_synchronized() 108 | infer_time = t2 - t1 109 | time_list[iii] = t2-t1 110 | 111 | # print('\n', len(pred), pred, recover) # list 长度是bs，代表每张图, 元素tensor，代表检测到的目标，每个tensor.shape [n, 6] xy4, conf, cls 112 | 113 | # Apply Classifier 114 | if classify: 115 | pred = apply_classifier(pred, modelc, img, im0s) 116 | 117 | # Process detections 118 | for i, det in enumerate(pred): # detections per image 119 | if opt.use_roi and det is not None: 120 | small_img_shape = torch.from_numpy(np.array([recover[1], recover[0]]).astype(np.float)) 121 | det[:, 0], det[:, 2] = det[:, 0] + recover[2], det[:, 2] + recover[2] 122 | det[:, 1], det[:, 3] = det[:, 1] + recover[3], det[:, 3] + recover[3] 123 | else: 124 | small_img_shape = img.shape[2::] 125 | if webcam: # batch_size >= 1 126 | p, s, im0 = path[i], '%g: ' % i, im0s[i].copy() 127 | else: 128 | p, s, im0 = path, '', im0s # im0s是原图 129 | 130 | save_path = str(Path(out) / Path(p).name) # output/filenamexxxx.jpg 131 | txt_path = str(Path(out) / Path(p).stem) + ('_%g' % dataset.frame if dataset.mode == 'video' else '') 132 | # output/filenamexxxx.txt 133 | s += '%gx%g ' % img.shape[2:] # print string, 640x640 134 | gn = torch.tensor(im0.shape)[[1, 0, 1, 0]] # normalization gain whwh 135 | # 本来是[720, 1280, 3]，重复取，变成[1280, 720, 1280, 720] 136 | if det is not None and len(det): 137 | # Rescale boxes from img_size to im0 size 138 | det[:, :4] = scale_coords(small_img_shape, det[:, :4], im0.shape).round() # 转换成原图的x1 y1 x2 y1，像素值 139 | # Print results 140 | for c in det[:, -1].unique(): 141 | n = (det[:, -1] == c).sum() # detections per class 142 | s += '%g %ss, ' % (n, names[int(c)]) # add to string # i.e. 1 crosswalk 143 | # s += f'{det[:, 4].item():.4f} ' 144 | # print(n) 145 | 146 | # Write results 147 | for *xyxy, conf, cls in det: 148 | if save_txt: # Write to file 149 | xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh 150 | with open(txt_path + '.txt', 'a') as f: 151 | x, y, w, h = xywh 152 | string = f"{int(cls)} {conf.item():.4f} {x:.6f} {y:.6f} {w:.6f} {h:.6f}\n" 153 | f.write(string) # label format 154 | 155 | if save_img or view_img: # Add bbox to image 156 | label = '%s %.2f' % (names[int(cls)], conf) 157 | # print(type(im0), im0.shape) array, 720, 1280, 3 158 | if names[int(cls)] in opt.plot_classes: 159 | # color = colors[int(cls)] 160 | color = (255, 85, 33) 161 | plot_one_box(xyxy, im0, label=label, color=color, line_thickness=5) 162 | 163 | # Print time (inference + NMS) 164 | prt_str = '%sDone. (%.5fs)' % (s, t2 - t1) 165 | # print(prt_str) 166 | os.system(f'echo "{prt_str}" >> {opt.output}/detect.log') 167 | 168 | # Stream results 169 | if view_img: 170 | cv2.imshow(p, im0) 171 | if cv2.waitKey(1) == ord('q'): # q to quit 172 | raise StopIteration 173 | 174 | # Save results (image with detections) 175 | if save_img: 176 | if dataset.mode == 'images': 177 | im0 = dp.dmpost(im0, det, det_id=detected_img_id, filename=Path(p).name, names=names) 178 | cv2.imwrite(save_path, im0) 179 | else: 180 | if vid_path != save_path: # new video 181 | vid_path = save_path 182 | if isinstance(vid_writer, cv2.VideoWriter): 183 | vid_writer.release() # release previous video writer 184 | 185 | fourcc = 'mp4v' # output video codec 186 | fps = vid_cap.get(cv2.CAP_PROP_FPS) 187 | w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH)) 188 | h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) 189 | vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*fourcc), fps, (w, h)) 190 | # print(detected_img_id, p, txt_path) 191 | tmp_filename = Path(txt_path).stem 192 | im0 = dp.dmpost(im0, det, det_id=detected_img_id, filename=tmp_filename, names=names) 193 | vid_writer.write(im0) 194 | detected_img_id += 1 195 | bar.set_description(f'inf_time: {infer_time*1000:.2f}ms {prt_str:<40}') 196 | 197 | if save_txt or save_img: 198 | print('Results saved to %s' % out) 199 | if platform == 'darwin' and not opt.update: # MacOS 200 | os.system('open ' + save_path) 201 | 202 | print('Done. (%.3fs)' % (time.time() - t0)) 203 | time_arr = np.array(time_list) 204 | prnt = f'Done. Network mean inference time: {np.mean(time_arr)*1000:.2f}ms, Mean FPS: {1/np.mean(time_arr):.2f}.' 205 | print(f'\nModel size {opt.img_size} inference {prnt}') 206 | os.system(f'echo "{prnt}" >> {opt.output}/detect.log') 207 | os.system(f'echo "useroi {opt.img_size} {prnt}" >> detect2.log') 208 | 209 | 210 | def run(opt, dp): 211 | with torch.no_grad(): 212 | if opt.update: # update all models (to fix SourceChangeWarning) 213 | for opt.weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']: 214 | detect() 215 | strip_optimizer(opt.weights) 216 | else: 217 | detect(opt, dp) 218 | 219 | 220 | def get_opt(): 221 | parser = argparse.ArgumentParser() 222 | parser.add_argument('--weights', nargs='+', type=str, required=False, 223 | default='runs/SEm_NST1_fog1_ep100/weights/best.pt', 224 | help='trained model path model.pt ddpath(s)') 225 | parser.add_argument('--source', type=str, default='example/images', help='source') # file/folder, 0 for webcam 226 | parser.add_argument('--output', type=str, default='example/output', help='output folder') # output folder 227 | parser.add_argument('--img-size', type=int, default=640, help='inference size (pixels)') 228 | parser.add_argument('--conf-thres', type=float, default=0.4, help='object confidence threshold') 229 | parser.add_argument('--iou-thres', type=float, default=0.5, help='IOU threshold for NMS') 230 | parser.add_argument('--device', default='0', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') 231 | parser.add_argument('--view-img', action='store_true', help='display results') 232 | parser.add_argument('--save-txt', type=bool, default=True, help='save results to *.txt') 233 | parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --class 0, or --class 0 2 3') 234 | parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS') 235 | parser.add_argument('--augment', action='store_true', help='augmented inference') 236 | parser.add_argument('--update', action='store_true', help='update all models') 237 | parser.add_argument('--control-line-setting', type=str, default='settings/cl_setting.yaml', 238 | help='control line setting') 239 | parser.add_argument('--select-control-line', type=str, default='general', 240 | help='select which control line. i.e. general, 0036') 241 | parser.add_argument('--field-size', type=int, default=5, help='receptive field size for post') 242 | parser.add_argument('--plot-classes', type=list, default=['crosswalk'], 243 | help='specifies which classes will be drawn') 244 | parser.add_argument('--not-use-ROI', action='store_true', 245 | help='not use roi for accelerate inference speed if there is the flag') 246 | parser.add_argument('--not-use-SSVM', action='store_true', 247 | help='not use ssvm method for analyse vehicle crossing behavior if there is the flag') 248 | 249 | opt = parser.parse_args() 250 | return opt 251 | 252 | 253 | if __name__ == '__main__': 254 | opt = get_opt() 255 | opt.weights = 'runs/m_ep300/weights/best.pt' 256 | opt.not_use_ROI = True 257 | opt.not_use_SSVM = True 258 | 259 | opt.use_roi = not opt.not_use_ROI 260 | opt.use_ssvm = not opt.not_use_SSVM 261 | for_paper = False 262 | if for_paper: 263 | exps = [ 264 | 'm_ep300', 'SEm_NST0_fog0_ep100', 'SEm_NST1_fog0_ep100', 265 | 'SEm_NST1_fog0_ep300', 'SEm_NST1_fog1_ep100'] 266 | exp = exps[1] 267 | 268 | roi = int(opt.use_roi) 269 | ssvm = int(opt.use_ssvm) 270 | 271 | opt.weights = f"runs/{exp}/weights/best.pt" 272 | opt.source = "/home/zzd/datasets/crosswalk/testsets_1770/Images" 273 | opt.output = f"/home/zzd/datasets/crosswalk/testsets_1770/{exp}_sz{opt.img_size}_ROI{roi}_SSVM{ssvm}" 274 | 275 | dp = post.DmPost(opt) 276 | 277 | print(opt) 278 | runtime = time.time() 279 | run(opt, dp) 280 | print(f'Total runtime: {time.time()-runtime:.5f}') 281 | 282 | 283 | -------------------------------------------------------------------------------- /docs/DATASETS.md: -------------------------------------------------------------------------------- 1 | # DATASETS 2 | 3 | The crosswalk detection datasets contains 3080 images for train and 1770 images for test. 4 | 5 | ## Trainset 6 | + Download from [Baidu Pan](https://pan.baidu.com/s/1jAsQ_kbU7cX0AqK4wDm6LA), passwd: **hul6** 7 | + Download from [Google Drive](https://drive.google.com/file/d/1RIIpdrAUXZRuoOgHMIr-BSsJxSjqiKC8/view?usp=sharing). 8 | + Download from [SJTU Jbox](https://jbox.sjtu.edu.cn/l/Z0i6nQ). 9 | + or use wget: 10 | ``` 11 | wget -c -O train_data_v5_format_3080.tar https://jbox.sjtu.edu.cn:10081/v2/delivery/data/de2ac2610893499686d095f174aa6ffd/?token= 12 | ``` 13 | **Notes** 14 | 15 | Please remove proxy if download failed. 16 | 17 | Unzip: 18 | ``` 19 | tar -xvf train_data_v5_format_3080.tar 20 | cd train_data_v5_format_3080 21 | rm labels/*.cache # if you are the first time to use this datasets. 22 | ``` 23 | The trainsets contains 3080 images, includes training and verification set, and the file structure format is YOLOv5 likes: 24 | ``` 25 | |--train_data_v5_format 26 | --|--images # the .jpg images 27 | --|--train # trainning images 28 | -- xxx.jpg 29 | -- ... 30 | --|--test # validation images 31 | --|--labels # corresponding .txt labels 32 | --|--train # trainning labels 33 | -- xxx.txt 34 | -- ... 35 | --|--test # validation labels 36 | ``` 37 | Each .txt file contains annotations in the format of CLS XC YC W H in each line. 38 | 39 | CLS(Classes): crosswalk, guide_arrows 40 | 41 | XC YC W H in terms of percentage. 42 | 43 | 44 | ## Testsets 45 | + Download from [Baidu Pan](https://pan.baidu.com/s/1-AWw2AjE8zTP-iVjOoifOg), passwd: **vj6b** 46 | + Download from [Google Drive](https://drive.google.com/file/d/1DBVt81LehEyiTxuUn-BUGBoh3gXcwrEE/view?usp=sharing). 47 | + Download from [SJTU Jbox](https://jbox.sjtu.edu.cn/l/eHE7vD). 48 | + or use wget: 49 | ``` 50 | wget -c -O testsets_1770.tar https://jbox.sjtu.edu.cn:10081/v2/delivery/data/a526b168356949068e6a13e8c254a71b/?token= 51 | ``` 52 | 53 | Unzip: 54 | ``` 55 | tar -xvf testsets_1770.tar 56 | cd testsets_1770 57 | ``` 58 | The testsets contains 1770 images. 59 | 60 | Currently we do not provide the testsets for vehicle crossing behavior analysis. 61 | 62 | 63 | 64 | -------------------------------------------------------------------------------- /docs/INSTALL.md: -------------------------------------------------------------------------------- 1 | # Installation 2 | 3 | This document describes how to get CNDet source code and configure the running environment. 4 | 5 | ## Requirements 6 | + Python 3.6+ 7 | + Pytorch 1.6+ 8 | + Linux or MacOS 9 | 10 | ## Get CDNet Code 11 | 12 | ##### use git: 13 | ```python 14 | git clone https://github.com/zhangzhengde0225/CDNet.git 15 | cd cdnet 16 | ``` 17 | ##### from github web: 18 | + Browse the [CDnet] repository 19 | + Click "Code"-"Download ZIP" 20 | ```python 21 | unzip CDNet-master.zip 22 | cd CDNet_master 23 | ``` 24 | 25 | 26 | ## Configure Environment 27 | [Anaconda](https://www.anaconda.com) is highly recommanded. 28 | 29 | Haven't Anaconda been installed yet? Download anaconda installer [here](https://www.anaconda.com/products/individual#Downloads) and install it: 30 | ```python 31 | chmod +x ./Anaconda3-2020.11-Linux-.sh 32 | ./Anaconda3-2020.11-Linux-.sh # install 33 | which conda # verify installation 34 | ``` 35 | 36 | After having a CONDA, directly import the fully configured environment: 37 | ```python 38 | conda env creat -f conda_cdnet_env.yaml 39 | ``` 40 | 41 | or Creating a environment from sratch: 42 | ```python 43 | conda create --name cdnet python=3.7 # create a env named as "cdnet" 44 | conda activate cdnet # activate the env 45 | which pip # verify pip 46 | pip install -r requirements.txt # install packages use pip 47 | # or use conda to install package 48 | conda install 49 | ``` 50 | 51 | 52 | 53 | 54 | -------------------------------------------------------------------------------- /docs/MODEL_ZOO.md: -------------------------------------------------------------------------------- 1 | # MODEL ZOO 2 | 3 | ## Pytorch model 4 | 5 | |Model|Detect Size|Speed (FPS)|F1 Score (%)|Pytorch Model|TensorRT Engine| 6 | |:---:|:---:|:---:|:---:|:---:|:---:| 7 | |YOLOv5m (baseline)|288x192|20.0|81.34|[Download](https://pan.baidu.com/s/1FCW5urZynjXjteR1E_2XuA) passwd: **duap**|[Download](link) 8 | |SE-YOLOv5m+NST+ROI+SSVM (CDNet)|288X192|33.1|94.72|[Download](https://pan.baidu.com/s/18p2TS-x830X7IUAVsdihow) passwd: **1cpp**|[Download](link) 9 | 10 | **Notes** 11 | 12 | Detection speed measured on **Jetson Nano** (4GB) with JetPack-4.4, cuda-10.2 and TensorRT-7.1.3. 13 | 14 | We also tested the GPU inference time on **RTX 3080** using [TensorRT](https://developer.nvidia.com/tensorRt), and the result shows that the speed is about 323 FPS (3.1 ms) per image. 15 | 16 | Note that TensorRT is optimized for specific TensorRT version, CUDA version and hardware, so the engine downloaded from this page can only be used on Jetson Nano with specific package installed. 17 | 18 | If you want to export the TensorRT engine from the pytorch weights for specific hardware and package, please refers to [here](https://github.com/wang-xinyu/tensorrtx). -------------------------------------------------------------------------------- /docs/Press_Material.md: -------------------------------------------------------------------------------- 1 | 2 | 今天带来一个基于改进YOLOv5的斑马线和汽车过线行为检测算法，文章提出多项tricks，数据集和复现代码开源！ 3 | 4 | 标题: CDNet: A Real-Time and Robust Crosswalk Detection Network on Jetson Nano Based on YOLOv5 5 | CDNet: 一个基于YOLOv5的在Jetson Nano上实时、鲁棒的斑马线检测网络 6 | 作者：Zheng-De Zhang, Meng-Lu Tan, Zhi-Cai Lan, Hai-Chun Liu, Ling Pei, Wen-Xian Yu 7 | 时间：Feb, 2022 8 | 期刊：Neural Computing & Applications, IF 5.6 9 | 10 | 图形摘要： 11 | 12 | 13 | 摘要： 14 | 在复杂场景和有限计算能力下实现实时、鲁棒的斑马线（人行横道）检测是当前智能交通管理系统（ITMS）的重要难点之一。有限的边缘计算能力和多云、晴天、雨天、雾天和夜间等真实复杂的场景同时对这项任务提出了挑战。本研究提出基于改进YOLOv5的人行横道检测网络（CDNet），实现车载摄像头视觉下快速准确的人行横道检测，并在Jetson nano设备上实现实时检测。强大的卷积神经网络特征提取器用于处理复杂环境，网络中嵌入了squeeze-and-excitation（SE）注意力机制模块，使用负样本训练（NST）方法提高准确率，利用感兴趣区域（ROI）算法进一步提高检测速度，提出了一种新的滑动感受野短时向量记忆（SSVM）算法来提高车辆交叉行为检测精度，使用合成雾增强算法允许模型适应有雾的场景。最后，在 Jetson nano 上以 33.1 FPS 的检测速度，我们在上述复杂场景中获得了 94.83% 的平均 F1 分数。对于晴天和阴天等更好的天气条件，F1 分数超过 98%。该工作为人工神经网络算法优化方法在边缘计算设备上的具体应用提供了参考，发布的模型为智能交通管理系统提供了算法支持。 15 | 16 | 贡献： 17 | + 注意力机制网络改进网络，提升精度，略微降低速度：SENet (Squeeze-and-Excitation Network) 18 | + 负样本训练，提升精度，速度不变: NST (Negative Samples Training) 19 | + 感兴趣区域，提升速度，精度下降：ROI (Region Of Interest) 20 | + 滑动感受野短时向量记忆算法，迁移斑马线检测任务到汽车过线行为检测任务，提升精度，速度不变：SSVM (Slide receptive field Short-term Vectors Memory) 21 | + 合成雾增强算法，增强数据集，适应雾天，提升精度，速度不变：SFA (Synthetic Fog Augment) 22 | + 斑马数据集：标注好的，车载摄像头视角下的，共计6868张图像。 23 | + 复现代码：见github项目主页 24 | 25 | 26 | 数据集、复现代码：https://github.com/zhangzhengde0225/CDNet 27 | 视频样例：https://www.bilibili.com/video/BV1qf4y1B7BA 28 | 阅读全文：https://rdcu.be/cHuc8 29 | 下载全文：https://doi.org/10.1007/s00521-022-07007-9 30 | 如果该项目对你有帮助，请点击github项目右上角star收藏和引用本论文，谢谢~ -------------------------------------------------------------------------------- /docs/README_zh_cn.md: -------------------------------------------------------------------------------- 1 | 2 | #### [English](https://github.com/zhangzhengde0225/CDNet) | 简体中文 3 | 4 | 如果本项目对你有帮助，请点击项目右上角star支持一下和引用下面的论文。 5 | 6 | # CDNet 7 | 8 | 这个项目是论文的《CDNet: 一个基于YOLOv5的在Jetson Nano上实时、鲁棒的斑马线检测网络》的代码、数据集和教程。 9 | 10 | CDNet (Crosswalk Detection Network) 是车载摄像头视野下检测斑马线（人行横道）和分析车辆过线行为的具体实现。 11 | 12 | ![GA](https://zhangzhengde0225.github.io/images/CDNet_GA.jpg) 13 | 14 | Fig.1 图形摘要 15 | 16 | # Highlights 17 | + 实现了斑马线检测和车辆过线行为检测。 18 | + 在特定任务中准确率和速度超过原生YOLOv5。 19 | + 在阴天、晴天、雨天、夜间等真实复杂场景中实现高鲁棒性。 20 | + 在 Jetson nano 边缘计算设备上实现实时检测 (33.1 FPS)。 21 | + 提供了标注好的斑马线数据集，共计6868张图 。 22 | 23 | # 贡献Contribution 24 | 25 | + 注意力机制网络改进网络，提升精度，略微降低速度：SENet (Squeeze-and-Excitation Network) 26 | + 负样本训练，提升精度，速度不变: NST (Negative Samples Training) 27 | + 感兴趣区域，提升速度，精度下降：ROI (Region Of Interest) 28 | + 滑动感受野短时向量记忆算法，迁移斑马线检测任务到汽车过线行为检测任务，提升精度，速度不变：SSVM (Slide receptive field Short-term Vectors Memory) 29 | + 合成雾增强算法，增强数据集，适应雾天，提升精度，速度不变：SFA (Synthetic Fog Augment) 30 | 31 | # 安装Installation 32 | 安装CDNet代码并配置环境，请查看： 33 | [docs/INSTALL.md](https://github.com/zhangzhengde0225/CDNet/blob/master/docs/INSTALL.md) 34 | 35 | # 模型库Model Zoo 36 | 模型库请查看： 37 | [docs/MODEL_ZOO.md](https://github.com/zhangzhengde0225/CDNet/blob/master/docs/MODEL_ZOO.md) 38 | 39 | # 数据集Datasets 40 | 下载训练集和测试集，请查看： 41 | [docs/DATASETS.md](https://github.com/zhangzhengde0225/CDNet/blob/master/docs/DATASETS.md) 42 | 43 | # 快速开始Quick Start 44 | ## 训练Train 45 | 46 | 安装CDNet代码，配置环境和下载数据集后，输入代码训练： 47 | ``` 48 | python train.py --trainset_path 49 | (such as: /home/xxx/datasets/train_data_yolov5_format) 50 | ``` 51 | 训练结果和权重将保存在 runs/xxx/ 目录中。 52 | 53 | 主要的可选参数： 54 | ``` 55 | --device "0, 1" # cpu or gpu id, "0, 1" means use two gpu to train. 56 | --img-size 640 57 | --batch-size 32 58 | --epochs 100 59 | --not-use-SE # use original YOLOv5 which not SE-module embedded if there is the flag 60 | ``` 61 | 62 | ## 推理Inference 63 | 逐图或视频检测斑马线并分析车辆过线行为，输入代码： 64 | ``` 65 | python detect.py 66 | ``` 67 | 68 | 主要的可选参数： 69 | ``` 70 | --source example/images # images dir 71 | --output example/output # output dir 72 | --img-size 640 # inference model size 73 | --device "0" # use cpu or gpu(gpu id) 74 | --plot-classes ["crosswalk"] # plot classes 75 | --field-size 5 # the Slide receptive field size of SSVM 76 | --not-use-ROI # not use roi for accelerate inference speed if there is the flag 77 | --not-use-SSVM # not use ssvm method for analyse vehicle crossing behavior if there is the flag 78 | ``` 79 | 80 | 关于安装和数据集的更多详情请参考： 81 | [docs/INSTALL.md](https://github.com/zhangzhengde0225/CDNet/blob/master/docs/INSTALL.md) and [docs/DATASETS.md](https://github.com/zhangzhengde0225/CDNet/blob/master/docs/DATASETS.md). 82 | 83 | ## 合成雾增强Fogging Augment 84 | 如果你想通过合成雾算法增加数据集，只需运行： 85 | ``` 86 | python fog_augment.py 87 | ``` 88 | 更多细节请查看fog_augment.py和/scripts/synthetic_fog.py中的源代码 89 | 90 | # 结果Results 91 | 92 | ![Results](https://zhangzhengde0225.github.io/images/CDNet_Results.jpg) 93 | Fig.2 与原生YOLOv5性能对比Performance compared to YOLOv5. 94 | 95 | **与原生YOLOv5相比，检测尺寸为640时，CDNet 在 Jetson nano 上提高了5.13%的F1分数和10.7FPS的速度，** 96 | **检测尺寸为288时，提升为13.38%的F1分数和13.1FPS。** 97 | 98 | # 贡献者Contributors 99 | CDNet的作者是: Zhengde Zhang, Menglu Tan, Zhicai Lan, Haichun Liu, Ling Pei and Wenxian Yu。 100 | 101 | 目前，CDNet由 102 | Zhengde Zhang (drivener@163.com)负责维护。 103 | 104 | 如果您有任何问题，请随时与我们联系。 105 | 106 | Zhengde Zhang的学术主页: [zhangzhengde0225.github.io](https://zhangzhengde0225.github.io). 107 | 108 | # 致谢Acknowledgement 109 | 110 | 本工作得到了国家自然科学基金会的支持[编号：61873163]。 111 | 112 | 我们非常感谢上海交通大学高性能计算中心提供计算资源。 113 | 114 | 我们非常感谢 115 | [yolov5](https://github.com/ultralytics/yolov5) 116 | 项目提供的目标检测算法基准。 117 | 118 | 我们非常感谢 119 | [tensorrtx](https://github.com/wang-xinyu/tensorrtx) 120 | 项目提供的模型模署到Jetson nano上的技术。 121 | 122 | # 链接Links 123 | 124 | B站视频样例：[https://www.bilibili.com/video/BV1qf4y1B7BA](https://www.bilibili.com/video/BV1qf4y1B7BA) 125 | 126 | 阅读论文全文：[https://rdcu.be/cHuc8](https://rdcu.be/cHuc8) 127 | 128 | 下载论文全文：[https://doi.org/10.1007/s00521-022-07007-9](https://doi.org/10.1007/s00521-022-07007-9) 129 | 130 | CSDN项目简介：[http://t.csdn.cn/Cf7c7](http://t.csdn.cn/Cf7c7) 131 | 132 | 如果对您有帮助，请为点击项目右上角的star支持一下或引用论文。 133 | 134 | # 引用Citation 135 | ``` 136 | @article{CDNet, 137 | author={Zheng-De Zhang, Meng-Lu Tan, Zhi-Cai Lan, Hai-Chun Liu, Ling Pei and Wen-Xian Yu}, 138 | title={CDNet: a real-time and robust crosswalk detection network on Jetson nano based on YOLOv5}, 139 | Journal={Neural Computing and Applications}, 140 | Year={2022}, 141 | DOI={10.1007/s00521-022-07007-9}, 142 | } 143 | ``` 144 | 145 | 146 | # 许可License 147 | CDNet可免费用于非商业用途，并可在这些条件下重新分发。 如需商业咨询，请发送电子邮件至 148 | drivener@163.com，我们会将详细协议发送给您。 149 | 150 | 151 | 152 | 153 | 154 | 155 | 156 | 157 | 158 | -------------------------------------------------------------------------------- /evaluate.py: -------------------------------------------------------------------------------- 1 | """ 2 | evaluate model performance on test set 3 | (1) 指定detect后的目标文件夹，获取评估结果 4 | (2) 指定不存在的检测文件夹，需要指定权重、ROI等一系列参数，获取评估结果 5 | """ 6 | import os 7 | import argparse 8 | from pathlib import Path 9 | import re 10 | 11 | import detect 12 | from scripts import post 13 | from scripts.get_confusing_matrix import GetConfusingMatrix 14 | 15 | 16 | class CDNetEvaluator(object): 17 | def __init__(self): 18 | self.gcm = GetConfusingMatrix() 19 | 20 | def __call__(self, opt): 21 | exps = sorted(os.listdir('./runs')) 22 | assert len(exps), f'Experiments not exists, please train model first' 23 | q = "\n".join([f'[{i:>2}] {x}' for i, x in enumerate(exps)]) 24 | ipt = input( 25 | f'Existing experiments:\n{q}\nPlease select which model to evaluate\n' 26 | f'(0 for model0, 0,1,2 for model0,1,2): ') 27 | ipt = '0' if ipt == '' else ipt 28 | indexes = [int(x) for x in ipt.split(',')] if ',' in ipt else [int(ipt)] 29 | for idx in indexes: 30 | print('\n') 31 | exp = exps[idx] 32 | opt.weights = f'./runs/{exp}/weights/best.pt' 33 | 34 | roi = int(opt.use_roi) 35 | ssvm = int(opt.use_ssvm) 36 | tp = f'{Path(opt.source).parent}/{exp}_sz{opt.img_size}_ROI{roi}_SSVM{ssvm}{opt.fog}' # target path 37 | 38 | flag = 0 # 0: detect and evaluate, 1: direct evaluate 39 | if os.path.exists(tp): 40 | ipt = input(f'Output_path: {tp} exists, \nevaluate without detect [YES/no]: ') 41 | if ipt in ['Yes', 'y', 'Y', 'yes', '']: 42 | flag = 1 43 | elif ipt in ['No', 'N', 'n']: 44 | flag = 0 45 | else: 46 | raise NotImplementedError(f'input error: {ipt}') 47 | if flag == 0: 48 | opt.output = tp 49 | print(f'opt: {opt}') 50 | dp = post.DmPost(opt) 51 | detect.run(opt, dp) 52 | 53 | ret = self.gcm.analyse(thresh=0.5, v5_out=tp, pn_dir=Path(opt.source).parent) 54 | time, ret = self.read_time_and_fps(tp=tp) 55 | 56 | def read_time_and_fps(self, tp): 57 | with open(f'{tp}/detect.log', 'r') as f: 58 | log = f.readlines() 59 | time, fps = re.findall(f'\d+.\d+', log[-1]) 60 | time = time[:-1] 61 | fps = fps[:-3] 62 | return time, fps 63 | 64 | def read_and_update_opt(eval_opt): 65 | det_opt = detect.get_opt() 66 | det_opt.source = eval_opt.source 67 | det_opt.output = eval_opt.output 68 | det_opt.not_use_ROI = eval_opt.not_use_ROI 69 | det_opt.not_use_SSVM = eval_opt.not_use_SSVM 70 | det_opt.img_size = eval_opt.img_size 71 | 72 | det_opt.use_roi = not det_opt.not_use_ROI 73 | det_opt.use_ssvm = not det_opt.not_use_SSVM 74 | return det_opt 75 | 76 | 77 | if __name__ == '__main__': 78 | parser = argparse.ArgumentParser() 79 | parser.add_argument('--source', type=str, default='example/images', help='source') 80 | parser.add_argument('--output', type=str, default='example/output', help='output folder') # output folder 81 | parser.add_argument('--img-size', type=int, default=640, help='inference size (pixels)') 82 | parser.add_argument('--not-use-ROI', action='store_true', 83 | help='not use roi for accelerate inference speed if there is the flag') 84 | parser.add_argument('--not-use-SSVM', action='store_true', 85 | help='not use ssvm method for analyse vehicle crossing behavior if there is the flag') 86 | opt = parser.parse_args() 87 | 88 | # opt.source = "/home/zzd/datasets/crosswalk/testsets_1770/Images" 89 | opt.source = "/home/zzd/datasets/crosswalk/testsets_1770/fogged_Images" 90 | opt = read_and_update_opt(opt) 91 | opt.fog = '_fogged' if 'fogged' in opt.source else '' 92 | 93 | CDNetE = CDNetEvaluator() 94 | CDNetE(opt) 95 | 96 | -------------------------------------------------------------------------------- /example/fog_image/fogged_br0.1_th0.05.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/fog_image/fogged_br0.1_th0.05.jpg -------------------------------------------------------------------------------- /example/fog_image/fogged_br0.3_th0.05.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/fog_image/fogged_br0.3_th0.05.jpg -------------------------------------------------------------------------------- /example/fog_image/fogged_br0.5_th0.05.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/fog_image/fogged_br0.5_th0.05.jpg -------------------------------------------------------------------------------- /example/fog_image/fogged_br0.7_th0.01.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/fog_image/fogged_br0.7_th0.01.jpg -------------------------------------------------------------------------------- /example/fog_image/fogged_br0.7_th0.03.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/fog_image/fogged_br0.7_th0.03.jpg -------------------------------------------------------------------------------- /example/fog_image/fogged_br0.7_th0.05.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/fog_image/fogged_br0.7_th0.05.jpg -------------------------------------------------------------------------------- /example/fog_image/fogged_br0.7_th0.07.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/fog_image/fogged_br0.7_th0.07.jpg -------------------------------------------------------------------------------- /example/fog_image/fogged_br0.7_th0.09.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/fog_image/fogged_br0.7_th0.09.jpg -------------------------------------------------------------------------------- /example/fog_image/fogged_br0.9_th0.05.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/fog_image/fogged_br0.9_th0.05.jpg -------------------------------------------------------------------------------- /example/fog_image/raw.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/fog_image/raw.jpg -------------------------------------------------------------------------------- /example/images/complex_scenarios/clearly_visible_0760_00001.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/complex_scenarios/clearly_visible_0760_00001.jpg -------------------------------------------------------------------------------- /example/images/complex_scenarios/damaged_0036_00846.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/complex_scenarios/damaged_0036_00846.jpg -------------------------------------------------------------------------------- /example/images/complex_scenarios/dazzling_0036_01707.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/complex_scenarios/dazzling_0036_01707.jpg -------------------------------------------------------------------------------- /example/images/complex_scenarios/distorted_0760_01846.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/complex_scenarios/distorted_0760_01846.jpg -------------------------------------------------------------------------------- /example/images/complex_scenarios/occluded_by_wiper_0760_02902.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/complex_scenarios/occluded_by_wiper_0760_02902.jpg -------------------------------------------------------------------------------- /example/images/complex_scenarios/partialy_lost_0036_01558.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/complex_scenarios/partialy_lost_0036_01558.jpg -------------------------------------------------------------------------------- /example/images/complex_scenarios/reflection_after_rain_0760_02711.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/complex_scenarios/reflection_after_rain_0760_02711.jpg -------------------------------------------------------------------------------- /example/images/complex_scenarios/truncated_in_steering_0760_03208.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/complex_scenarios/truncated_in_steering_0760_03208.jpg -------------------------------------------------------------------------------- /example/images/complex_scenarios/view_blocked_0036_05420.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/complex_scenarios/view_blocked_0036_05420.jpg -------------------------------------------------------------------------------- /example/images/complex_scenarios/visible_at_night_0036_01127.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/complex_scenarios/visible_at_night_0036_01127.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00030.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00030.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00031.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00031.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00032.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00032.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00033.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00033.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00034.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00034.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00035.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00035.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00036.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00036.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00037.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00037.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00038.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00038.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00039.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00039.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00040.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00040.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00041.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00041.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00042.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00042.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00043.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00043.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00044.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00044.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00045.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00045.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00046.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00046.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00047.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00047.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00048.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00048.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00049.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00049.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00050.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00050.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00051.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00051.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00052.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00052.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00053.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00053.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00054.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00054.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00055.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00055.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00056.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00056.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00057.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00057.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00058.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00058.jpg -------------------------------------------------------------------------------- /example/images/exmp_imgs/filename_00059.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/exmp_imgs/filename_00059.jpg -------------------------------------------------------------------------------- /example/images/filename_00030.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00030.jpg -------------------------------------------------------------------------------- /example/images/filename_00031.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00031.jpg -------------------------------------------------------------------------------- /example/images/filename_00032.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00032.jpg -------------------------------------------------------------------------------- /example/images/filename_00033.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00033.jpg -------------------------------------------------------------------------------- /example/images/filename_00034.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00034.jpg -------------------------------------------------------------------------------- /example/images/filename_00035.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00035.jpg -------------------------------------------------------------------------------- /example/images/filename_00036.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00036.jpg -------------------------------------------------------------------------------- /example/images/filename_00037.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00037.jpg -------------------------------------------------------------------------------- /example/images/filename_00038.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00038.jpg -------------------------------------------------------------------------------- /example/images/filename_00039.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00039.jpg -------------------------------------------------------------------------------- /example/images/filename_00040.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00040.jpg -------------------------------------------------------------------------------- /example/images/filename_00041.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00041.jpg -------------------------------------------------------------------------------- /example/images/filename_00042.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00042.jpg -------------------------------------------------------------------------------- /example/images/filename_00043.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00043.jpg -------------------------------------------------------------------------------- /example/images/filename_00044.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00044.jpg -------------------------------------------------------------------------------- /example/images/filename_00045.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00045.jpg -------------------------------------------------------------------------------- /example/images/filename_00046.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00046.jpg -------------------------------------------------------------------------------- /example/images/filename_00047.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00047.jpg -------------------------------------------------------------------------------- /example/images/filename_00048.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00048.jpg -------------------------------------------------------------------------------- /example/images/filename_00049.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00049.jpg -------------------------------------------------------------------------------- /example/images/filename_00050.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00050.jpg -------------------------------------------------------------------------------- /example/images/filename_00051.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00051.jpg -------------------------------------------------------------------------------- /example/images/filename_00052.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00052.jpg -------------------------------------------------------------------------------- /example/images/filename_00053.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00053.jpg -------------------------------------------------------------------------------- /example/images/filename_00054.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00054.jpg -------------------------------------------------------------------------------- /example/images/filename_00055.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00055.jpg -------------------------------------------------------------------------------- /example/images/filename_00056.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00056.jpg -------------------------------------------------------------------------------- /example/images/filename_00057.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00057.jpg -------------------------------------------------------------------------------- /example/images/filename_00058.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00058.jpg -------------------------------------------------------------------------------- /example/images/filename_00059.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/filename_00059.jpg -------------------------------------------------------------------------------- /example/images/for_overall_architecture/filename1120.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/for_overall_architecture/filename1120.jpg -------------------------------------------------------------------------------- /example/images/for_overall_architecture/filename1121.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/for_overall_architecture/filename1121.jpg -------------------------------------------------------------------------------- /example/images/for_overall_architecture/filename1122.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/for_overall_architecture/filename1122.jpg -------------------------------------------------------------------------------- /example/images/for_overall_architecture/filename1123.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/for_overall_architecture/filename1123.jpg -------------------------------------------------------------------------------- /example/images/for_overall_architecture/filename1124.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/for_overall_architecture/filename1124.jpg -------------------------------------------------------------------------------- /example/images/for_overall_architecture/filename1125.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/for_overall_architecture/filename1125.jpg -------------------------------------------------------------------------------- /example/images/for_overall_architecture/filename1126.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/for_overall_architecture/filename1126.jpg -------------------------------------------------------------------------------- /example/images/for_overall_architecture/filename1127.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/for_overall_architecture/filename1127.jpg -------------------------------------------------------------------------------- /example/images/for_overall_architecture/filename1128.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/for_overall_architecture/filename1128.jpg -------------------------------------------------------------------------------- /example/images/for_overall_architecture/filename1129.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/for_overall_architecture/filename1129.jpg -------------------------------------------------------------------------------- /example/images/for_overall_architecture/filename1130.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/for_overall_architecture/filename1130.jpg -------------------------------------------------------------------------------- /example/images/for_overall_architecture/filename1131.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/images/for_overall_architecture/filename1131.jpg -------------------------------------------------------------------------------- /example/out_for_overall/detect.log: -------------------------------------------------------------------------------- 1 | 160x640 1 crosswalks, Done. (0.323s) 2 | 160x640 1 crosswalks, Done. (0.285s) 3 | 160x640 1 guide_arrowss, Done. (0.275s) 4 | 160x640 1 guide_arrowss, Done. (0.281s) 5 | 160x640 1 guide_arrowss, Done. (0.279s) 6 | 160x640 1 crosswalks, 1 guide_arrowss, Done. (0.284s) 7 | 160x640 1 crosswalks, 1 guide_arrowss, Done. (0.275s) 8 | 160x640 1 crosswalks, 1 guide_arrowss, Done. (0.276s) 9 | 160x640 1 crosswalks, Done. (0.286s) 10 | 160x640 1 crosswalks, Done. (0.282s) 11 | 160x640 1 crosswalks, Done. (0.278s) 12 | 160x640 1 crosswalks, Done. (0.279s) 13 | -------------------------------------------------------------------------------- /example/out_for_overall/filename1120.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/out_for_overall/filename1120.jpg -------------------------------------------------------------------------------- /example/out_for_overall/filename1120.txt: -------------------------------------------------------------------------------- 1 | 0 0.7516 0.305469 0.609722 0.164062 0.019444 2 | -------------------------------------------------------------------------------- /example/out_for_overall/filename1121.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/out_for_overall/filename1121.jpg -------------------------------------------------------------------------------- /example/out_for_overall/filename1121.txt: -------------------------------------------------------------------------------- 1 | 0 0.5696 0.291797 0.609722 0.188281 0.019444 2 | -------------------------------------------------------------------------------- /example/out_for_overall/filename1122.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/out_for_overall/filename1122.jpg -------------------------------------------------------------------------------- /example/out_for_overall/filename1122.txt: -------------------------------------------------------------------------------- 1 | 1 0.6779 0.257031 0.644444 0.253125 0.030556 2 | -------------------------------------------------------------------------------- /example/out_for_overall/filename1123.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/out_for_overall/filename1123.jpg -------------------------------------------------------------------------------- /example/out_for_overall/filename1123.txt: -------------------------------------------------------------------------------- 1 | 1 0.8535 0.220313 0.671528 0.317187 0.040278 2 | -------------------------------------------------------------------------------- /example/out_for_overall/filename1124.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/out_for_overall/filename1124.jpg -------------------------------------------------------------------------------- /example/out_for_overall/filename1124.txt: -------------------------------------------------------------------------------- 1 | 1 0.9064 0.201562 0.716667 0.403125 0.086111 2 | -------------------------------------------------------------------------------- /example/out_for_overall/filename1125.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/out_for_overall/filename1125.jpg -------------------------------------------------------------------------------- /example/out_for_overall/filename1125.txt: -------------------------------------------------------------------------------- 1 | 1 0.9220 0.208594 0.761111 0.417188 0.150000 2 | 0 0.6014 0.218750 0.645139 0.381250 0.037500 3 | -------------------------------------------------------------------------------- /example/out_for_overall/filename1126.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/out_for_overall/filename1126.jpg -------------------------------------------------------------------------------- /example/out_for_overall/filename1126.txt: -------------------------------------------------------------------------------- 1 | 1 0.9169 0.220313 0.834722 0.440625 0.227778 2 | 0 0.8203 0.229297 0.656250 0.430469 0.051389 3 | -------------------------------------------------------------------------------- /example/out_for_overall/filename1127.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/out_for_overall/filename1127.jpg -------------------------------------------------------------------------------- /example/out_for_overall/filename1127.txt: -------------------------------------------------------------------------------- 1 | 0 0.8860 0.241016 0.665278 0.466406 0.066667 2 | 1 0.8282 0.239063 0.851389 0.478125 0.208333 3 | -------------------------------------------------------------------------------- /example/out_for_overall/filename1128.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/out_for_overall/filename1128.jpg -------------------------------------------------------------------------------- /example/out_for_overall/filename1128.txt: -------------------------------------------------------------------------------- 1 | 0 0.8918 0.248047 0.689583 0.496094 0.087500 2 | -------------------------------------------------------------------------------- /example/out_for_overall/filename1129.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/out_for_overall/filename1129.jpg -------------------------------------------------------------------------------- /example/out_for_overall/filename1129.txt: -------------------------------------------------------------------------------- 1 | 0 0.9231 0.273047 0.729167 0.539844 0.122222 2 | -------------------------------------------------------------------------------- /example/out_for_overall/filename1130.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/out_for_overall/filename1130.jpg -------------------------------------------------------------------------------- /example/out_for_overall/filename1130.txt: -------------------------------------------------------------------------------- 1 | 0 0.9285 0.314844 0.790278 0.629687 0.216667 2 | -------------------------------------------------------------------------------- /example/out_for_overall/filename1131.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/out_for_overall/filename1131.jpg -------------------------------------------------------------------------------- /example/out_for_overall/filename1131.txt: -------------------------------------------------------------------------------- 1 | 0 0.8945 0.471484 0.830556 0.942969 0.255556 2 | -------------------------------------------------------------------------------- /example/output/detect.log: -------------------------------------------------------------------------------- 1 | 160x640 Done. (0.306s) 2 | filename_00030.jpg detect_id: 0 speed: None count: 0 status: No crosswalk 3 | 160x640 Done. (0.283s) 4 | filename_00031.jpg detect_id: 1 speed: None count: 0 status: No crosswalk 5 | 160x640 Done. (0.279s) 6 | filename_00032.jpg detect_id: 2 speed: None count: 0 status: No crosswalk 7 | 160x640 Done. (0.290s) 8 | filename_00033.jpg detect_id: 3 speed: None count: 0 status: No crosswalk 9 | 160x640 Done. (0.284s) 10 | filename_00034.jpg detect_id: 4 speed: None count: 0 status: No crosswalk 11 | 160x640 Done. (0.276s) 12 | filename_00035.jpg detect_id: 5 speed: None count: 0 status: No crosswalk 13 | 160x640 1 guide_arrowss, Done. (0.281s) 14 | filename_00036.jpg detect_id: 6 speed: None count: 0 status: No crosswalk 15 | 160x640 1 crosswalks, 1 guide_arrowss, Done. (0.276s) 16 | filename_00037.jpg detect_id: 7 speed: None count: 0 status: No crosswalk 17 | 160x640 1 crosswalks, Done. (0.291s) 18 | filename_00038.jpg detect_id: 8 speed: 19.0 count: 0 status: Crossing 19 | 160x640 1 crosswalks, Done. (0.296s) 20 | filename_00039.jpg detect_id: 9 speed: 40.0 count: 0 status: Crossing 21 | 160x640 1 crosswalks, Done. (0.285s) 22 | 23 | filename_00040.jpg count+1, conf: 0.96 count: 1 24 | filename_00040.jpg detect_id: 10 speed: 36.0 count: 1 status: Crossing 25 | 160x640 1 crosswalks, Done. (0.285s) 26 | filename_00041.jpg detect_id: 11 speed: 40.0 count: 1 status: Crossing 27 | 160x640 1 crosswalks, Done. (0.284s) 28 | filename_00042.jpg detect_id: 12 speed: 16.0 count: 1 status: Crossing 29 | 160x640 1 crosswalks, Done. (0.278s) 30 | filename_00043.jpg detect_id: 13 speed: 35.0 count: 1 status: Crossing 31 | 160x640 Done. (0.284s) 32 | filename_00044.jpg detect_id: 14 speed: None count: 1 status: Crossing 33 | 160x640 Done. (0.286s) 34 | filename_00045.jpg detect_id: 15 speed: None count: 1 status: Crossing 35 | 160x640 1 crosswalks, Done. (0.278s) 36 | filename_00046.jpg detect_id: 16 speed: -66.33333333333333 count: 1 status: No crosswalk 37 | 160x640 2 crosswalks, Done. (0.282s) 38 | filename_00047.jpg detect_id: 17 speed: 12.0 count: 1 status: Crossing 39 | 160x640 1 crosswalks, Done. (0.280s) 40 | filename_00048.jpg detect_id: 18 speed: 14.0 count: 1 status: Crossing 41 | 160x640 1 crosswalks, Done. (0.283s) 42 | filename_00049.jpg detect_id: 19 speed: 41.0 count: 1 status: Crossing 43 | 160x640 1 crosswalks, Done. (0.291s) 44 | filename_00050.jpg detect_id: 20 speed: 85.0 count: 1 status: Crossing 45 | 160x640 Done. (0.277s) 46 | filename_00051.jpg detect_id: 21 speed: None count: 1 status: Crossing 47 | 160x640 Done. (0.283s) 48 | filename_00052.jpg detect_id: 22 speed: None count: 1 status: Crossing 49 | 160x640 Done. (0.285s) 50 | filename_00053.jpg detect_id: 23 speed: None count: 1 status: Crossing 51 | 160x640 Done. (0.276s) 52 | filename_00054.jpg detect_id: 24 speed: None count: 1 status: Crossing 53 | 160x640 Done. (0.282s) 54 | filename_00055.jpg detect_id: 25 speed: None count: 1 status: No crosswalk 55 | 160x640 Done. (0.286s) 56 | filename_00056.jpg detect_id: 26 speed: None count: 1 status: No crosswalk 57 | 160x640 Done. (0.278s) 58 | filename_00057.jpg detect_id: 27 speed: None count: 1 status: No crosswalk 59 | 160x640 Done. (0.286s) 60 | filename_00058.jpg detect_id: 28 speed: None count: 1 status: No crosswalk 61 | 160x640 Done. (0.287s) 62 | filename_00059.jpg detect_id: 29 speed: None count: 1 status: No crosswalk 63 | 0.28381s 3.5234FPS 64 | -------------------------------------------------------------------------------- /example/output/filename_00030.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00030.jpg -------------------------------------------------------------------------------- /example/output/filename_00031.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00031.jpg -------------------------------------------------------------------------------- /example/output/filename_00032.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00032.jpg -------------------------------------------------------------------------------- /example/output/filename_00033.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00033.jpg -------------------------------------------------------------------------------- /example/output/filename_00034.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00034.jpg -------------------------------------------------------------------------------- /example/output/filename_00035.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00035.jpg -------------------------------------------------------------------------------- /example/output/filename_00036.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00036.jpg -------------------------------------------------------------------------------- /example/output/filename_00036.txt: -------------------------------------------------------------------------------- 1 | 1 0.8876 0.275391 0.675000 0.271094 0.061111 2 | -------------------------------------------------------------------------------- /example/output/filename_00037.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00037.jpg -------------------------------------------------------------------------------- /example/output/filename_00037.txt: -------------------------------------------------------------------------------- 1 | 1 0.9397 0.233203 0.803472 0.466406 0.204167 2 | 0 0.5480 0.316016 0.637500 0.289844 0.038889 3 | -------------------------------------------------------------------------------- /example/output/filename_00038.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00038.jpg -------------------------------------------------------------------------------- /example/output/filename_00038.txt: -------------------------------------------------------------------------------- 1 | 0 0.8900 0.309766 0.663889 0.399219 0.055556 2 | -------------------------------------------------------------------------------- /example/output/filename_00039.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00039.jpg -------------------------------------------------------------------------------- /example/output/filename_00039.txt: -------------------------------------------------------------------------------- 1 | 0 0.9498 0.295312 0.719444 0.585938 0.108333 2 | -------------------------------------------------------------------------------- /example/output/filename_00040.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00040.jpg -------------------------------------------------------------------------------- /example/output/filename_00040.txt: -------------------------------------------------------------------------------- 1 | 0 0.9563 0.336719 0.770139 0.673437 0.173611 2 | -------------------------------------------------------------------------------- /example/output/filename_00041.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00041.jpg -------------------------------------------------------------------------------- /example/output/filename_00041.txt: -------------------------------------------------------------------------------- 1 | 0 0.9516 0.410156 0.825000 0.820312 0.233333 2 | -------------------------------------------------------------------------------- /example/output/filename_00042.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00042.jpg -------------------------------------------------------------------------------- /example/output/filename_00042.txt: -------------------------------------------------------------------------------- 1 | 0 0.9410 0.412891 0.847222 0.825781 0.191667 2 | -------------------------------------------------------------------------------- /example/output/filename_00043.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00043.jpg -------------------------------------------------------------------------------- /example/output/filename_00043.txt: -------------------------------------------------------------------------------- 1 | 0 0.8974 0.420312 0.896528 0.806250 0.112500 2 | -------------------------------------------------------------------------------- /example/output/filename_00044.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00044.jpg -------------------------------------------------------------------------------- /example/output/filename_00045.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00045.jpg -------------------------------------------------------------------------------- /example/output/filename_00046.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00046.jpg -------------------------------------------------------------------------------- /example/output/filename_00046.txt: -------------------------------------------------------------------------------- 1 | 0 0.5608 0.361328 0.619444 0.247656 0.033333 2 | -------------------------------------------------------------------------------- /example/output/filename_00047.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00047.jpg -------------------------------------------------------------------------------- /example/output/filename_00047.txt: -------------------------------------------------------------------------------- 1 | 0 0.7835 0.344531 0.636111 0.329688 0.038889 2 | 0 0.5951 0.113672 0.765278 0.227344 0.202778 3 | -------------------------------------------------------------------------------- /example/output/filename_00048.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00048.jpg -------------------------------------------------------------------------------- /example/output/filename_00048.txt: -------------------------------------------------------------------------------- 1 | 0 0.6838 0.316797 0.655556 0.432031 0.063889 2 | -------------------------------------------------------------------------------- /example/output/filename_00049.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00049.jpg -------------------------------------------------------------------------------- /example/output/filename_00049.txt: -------------------------------------------------------------------------------- 1 | 0 0.9292 0.311328 0.713194 0.614844 0.106944 2 | -------------------------------------------------------------------------------- /example/output/filename_00050.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00050.jpg -------------------------------------------------------------------------------- /example/output/filename_00050.txt: -------------------------------------------------------------------------------- 1 | 0 0.9086 0.427734 0.830556 0.849219 0.222222 2 | -------------------------------------------------------------------------------- /example/output/filename_00051.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00051.jpg -------------------------------------------------------------------------------- /example/output/filename_00052.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00052.jpg -------------------------------------------------------------------------------- /example/output/filename_00053.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00053.jpg -------------------------------------------------------------------------------- /example/output/filename_00054.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00054.jpg -------------------------------------------------------------------------------- /example/output/filename_00055.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00055.jpg -------------------------------------------------------------------------------- /example/output/filename_00056.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00056.jpg -------------------------------------------------------------------------------- /example/output/filename_00057.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00057.jpg -------------------------------------------------------------------------------- /example/output/filename_00058.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00058.jpg -------------------------------------------------------------------------------- /example/output/filename_00059.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/example/output/filename_00059.jpg -------------------------------------------------------------------------------- /fog_augment.py: -------------------------------------------------------------------------------- 1 | """ 2 | fogging train and test datasets using synthetic fog algorithm 3 | """ 4 | 5 | import os, sys 6 | import shutil 7 | from pathlib import Path 8 | import numpy as np 9 | from tqdm import tqdm 10 | import cv2 11 | import random 12 | from copy import deepcopy 13 | 14 | from scripts.synthetic_fog import SyntheticFog 15 | 16 | 17 | class AugmentCrosswalkDataset(object): 18 | def __init__(self, source_path): 19 | self.sp = source_path # source path 20 | p = Path(self.sp) 21 | self.tp = f'{p.parent}/fogged_{p.stem}' # target path 22 | 23 | self.sf = SyntheticFog() # synthetic fog object 24 | 25 | def augment(self, show=False): 26 | """augment train and test set in YOLOv5 format""" 27 | # 逐张进行增强 28 | sp = self.sp 29 | tp = self.tp 30 | print(f'fogged data will be saved to: {tp}') 31 | if os.path.exists(self.tp): 32 | 33 | shutil.rmtree(self.tp) 34 | os.makedirs(f'{self.tp}/images/train') 35 | os.makedirs(f'{self.tp}/images/test') 36 | os.makedirs(f'{self.tp}/labels/train') 37 | os.makedirs(f'{self.tp}/labels/test') 38 | 39 | for trte in ['train', 'test']: 40 | pi = f'{sp}/images/{trte}' # path of images 41 | pl = f'{sp}/labels/{trte}' 42 | ti = f'{tp}/images/{trte}' 43 | tl = f'{tp}/labels/{trte}' 44 | 45 | imgs = [f'{x}' for x in os.listdir(pi) if x.endswith('.jpg')] 46 | print(f'transform {trte} images, total: {len(imgs)}, transformed total: {2*len(img)}.') 47 | bar = tqdm(imgs) 48 | for i, img_name in enumerate(bar): 49 | img_path = f'{pi}/{img_name}' 50 | stem = Path(img_path).stem 51 | assert os.path.exists(img_path), f'img does not exists {img_path}' 52 | 53 | # 先拷贝原始图像和标注 54 | shutil.copy(img_path, f'{ti}/{img_name}') 55 | shutil.copy(f'{pl}/{stem}.txt', f'{tl}/{stem}.txt') 56 | 57 | # fogging 58 | img = cv2.imread(img_path) 59 | h, w, c = img.shape 60 | # random brightness and thickness 61 | br = np.clip(0.2 * np.random.randn() + 0.5, 0.1, 0.9) # 0.1~0.9 62 | th = np.clip(0.01 * np.random.randn() + 0.05, 0.01, 0.09) 63 | normed_img = img.copy()/255.0 64 | fogged_img = self.sf.fogging_img( 65 | normed_img, brightness=br, thickness=th, high_efficiency=True) 66 | fogged_img = np.array(fogged_img*255, dtype=np.uint8) 67 | 68 | # save fogged images and labels 69 | cv2.imwrite(f'{ti}/fogged_{img_name}', fogged_img) 70 | shutil.copy(f'{pl}/{stem}.txt', f'{tl}/fogged_{stem}.txt') 71 | 72 | if show: 73 | print(f'img_name: {Path(img_path).name} img: {img.shape} br: {br} th: {th} max: {np.max(fogged_img)}') 74 | self.show(img, name='src_img', wait=False) 75 | self.show(fogged_img, name='fogged_img', wait=False) 76 | if cv2.waitKey(0) == ord('q'): 77 | break 78 | 79 | bar.set_description(f'Img and fogged img saved, {stem}.') 80 | 81 | def show(self, img, name='xx', wait=True): 82 | h, w, c = img.shape 83 | scale = 0.5 84 | show_img = cv2.resize(img, (int(w*scale), int(h*scale))) 85 | cv2.imshow(name, show_img) 86 | if wait: 87 | cv2.waitKey(0) 88 | 89 | def augment_testset(self, dir): 90 | """augment only test set""" 91 | self.sf.fogging_dir(sp=dir, tp=None, random_params=True, save_src_img=True) 92 | 93 | 94 | if __name__ == '__main__': 95 | source_path = '/home/zzd/datasets/crosswalk/train_data_v5_format' 96 | acd = AugmentCrosswalkDataset(source_path) 97 | acd.augment(show=False) 98 | # test_imgs_path = '/home/zzd/datasets/crosswalk/testsets_1770/Images' 99 | # acd.augment_testset(test_imgs_path) 100 | 101 | 102 | 103 | 104 | 105 | 106 | 107 | 108 | 109 | 110 | 111 | -------------------------------------------------------------------------------- /for_paper/CDNet_plot.py: -------------------------------------------------------------------------------- 1 | """ 2 | 绘制result的图 3 | """ 4 | 5 | import os 6 | import matplotlib.pyplot as plt 7 | import yaml 8 | import numpy as np 9 | import matplotlib 10 | 11 | matplotlib.rcParams['xtick.direction'] = 'in' # 刻度线向内 12 | matplotlib.rcParams['ytick.direction'] = 'in' 13 | c = [ 14 | 'lightcoral', 'limegreen', 'royalblue', 'orange', 'mediumorchid', 15 | 'orangered', 'lawngreen', 'deepskyblue', 'gold', 'fuchsia', 16 | 'sandybrown', 'cyan'] 17 | c = [ 18 | 'lightcoral', 'limegreen', 'royalblue', 'orange', 'mediumorchid', 19 | 'deepskyblue', 'gold', 'fuchsia', 20 | 'sandybrown', 'cyan'] 21 | c = [ 22 | 'black', 'lightcoral', 'limegreen', 'royalblue', 'orange', 'mediumorchid', 23 | 'deepskyblue', '#EE3B3B', 'gold', 'fuchsia', 24 | 'sandybrown', 'cyan'] 25 | c = [ 26 | 'black', 'lightcoral', 'limegreen', 'royalblue', 'orange', 'mediumorchid', 27 | 'deepskyblue', '#009966', '#EE3B3B', 'gold', 'fuchsia', 28 | 'sandybrown', 'cyan'] 29 | 30 | 31 | class CDNetPlot(object): 32 | def __init__(self): 33 | self.data = self.load_data('data/results_3.yaml') 34 | 35 | def load_data(self, path): 36 | with open(path, 'r') as f: 37 | data = yaml.load(f, Loader=yaml.FullLoader) 38 | return data 39 | 40 | def plot(self): 41 | d = self.data 42 | for i, k in enumerate(d.keys()): 43 | v = [float(x) for x in d[k].split()] 44 | print(k, v, type(v)) 45 | x = np.array([v[0], v[2]]).astype(np.float32) 46 | y = np.array([v[1], v[3]]).astype(np.float32) 47 | # 划线 48 | lb = '$\mathbf{'+k+'}$' 49 | plt.plot( 50 | x, y, ls='-', label=lb, c=c[i], linewidth=3, 51 | marker='', markersize=10) 52 | # 画点288 53 | plt.plot(v[0], v[1], marker='^', c=c[i], markersize=15) 54 | plt.plot(v[2], v[3], marker='s', c=c[i], markersize=15) 55 | # 添加额外的标签 56 | if False: 57 | if i == 0: 58 | plt.text(v[0]-12, v[1]-1.5, '288', size=16) 59 | plt.text(v[2] - 12, v[3] - 1.5, '640', size=16) 60 | 61 | fontdict = {'size': 18, 'weight': 'bold'} 62 | plt.legend(loc=2, bbox_to_anchor=(1.0, 1.0), borderaxespad=0.) 63 | plt.grid() 64 | plt.xlabel('Detection Speed (FPS)', fontdict=fontdict) 65 | plt.ylabel('F1 score (%)', fontdict=fontdict) 66 | plt.xticks(range(0, 100, 5), size=12, weight='bold') 67 | plt.xlim(1, 42) 68 | plt.yticks(range(60, 101, 5), size=12, weight='bold') 69 | plt.ylim(78, 98) 70 | # plt.xticks(range(0, 300, 50)) 71 | plt.show() 72 | # plt.savefig('results.png', dpi=2000, bbox_inches='tight') 73 | 74 | 75 | if __name__ == '__main__': 76 | cdnetp = CDNetPlot() 77 | cdnetp.plot() 78 | -------------------------------------------------------------------------------- /for_paper/data/results.yaml: -------------------------------------------------------------------------------- 1 | Base (YOLOv5m): 189 81.34 634 90.34 2 | Base+NST: 189 89.03 634 91.18 3 | Base+NST+SE: 205 86.41 650 92.10 4 | Base+NST+ROI: 74 81.27 273 87.17 5 | Base+NST+SE+ROI: 86 79.33 284 86.25 6 | Base+NST+SE+SSVM: 205 94.86 650 95.47 7 | Base+NST+ROI+SSVM: 74 93.70 273 95.31 8 | Base+NST+SE+ROI+SSVM (CDNet): 86 94.72 284 95.47 -------------------------------------------------------------------------------- /for_paper/entropy_analyse.py: -------------------------------------------------------------------------------- 1 | """ 2 | 分析使用NST前后的数据集，根据信息论，从信息熵出发分析性能。 3 | """ 4 | import os 5 | import numpy as np 6 | 7 | 8 | class EntropyAnalyse(object): 9 | def __init__(self): 10 | self.lb_2cls = "/home/zzd/datasets/crosswalk/train_data_v5_format/labels/train" 11 | self.lb_1cls = "/home/zzd/datasets/crosswalk/train_data_v5_format_1cls/labels/train" 12 | 13 | self.out_1cls_288 = "/home/zzd/datasets/crosswalk/testsets_1770/dense_testsize_full_1cls_640_288_output" 14 | self.out_1cls_640 = "/home/zzd/datasets/crosswalk/testsets_1770/dense_testsize_full_1cls_640_640_output" 15 | self.lb_2cls_te = "/home/zzd/datasets/crosswalk/train_data_v5_format/labels/test" 16 | 17 | def run(self): 18 | 19 | e = self.get_error_rate(out_path=self.out_1cls_640) 20 | # print(e) 21 | # exit() 22 | # 第一步，读取概率 23 | P2cls = self.get_P_from_label(path=self.lb_2cls) 24 | H2 = self.cal_information_entropy(P2cls) 25 | print(H2) 26 | exit() 27 | P1cls = self.get_P_from_label(path=self.lb_1cls) 28 | 29 | 30 | # P = np.array([0.1, 0.9, 0]) 31 | H1 = self.cal_information_entropy(P1cls) # 熵 32 | 33 | print(H1, H2) 34 | 35 | # 统计在1cls中误检为crosswalk的数目，分288和640 36 | 37 | 38 | 39 | def get_error_rate(self, out_path): 40 | """ 41 | 根据2cls的测试集的标注文件和1cls时的检测结果，统计在所有标注的guide_arrows中，被误检为crosswalk的错误数 42 | :return: 43 | """ 44 | lb2te = self.lb_2cls_te 45 | txts = [x for x in os.listdir(lb2te) if x.endswith('.txt')] 46 | error = np.zeros(3) # 总标注数目，标注了1个同时检测为crosswalk，标注2个同时检测出2个crosswalk 47 | for i, txt in enumerate(txts): 48 | # print(txt) 49 | classes = self.read_txt(f'{lb2te}/{txt}') 50 | if '1' in classes: 51 | classes_ga = [x for x in classes if x == '1'] # 只有guide arrows 52 | both = len(classes) != len(classes_ga) # both为真时，表示该标注文件同时标注了crosswalk和guidearrows 53 | assert len(classes_ga) == 1 54 | classes2 = self.read_txt(f'{out_path}/{txt}') 55 | classes2 = [] if classes2 is None else classes2 56 | # print(classes, classes2, both) 57 | error[0] += 1 58 | 59 | if not both and '0' in classes2: 60 | error[1] += 1 61 | if both and len(classes2) >= 2: 62 | print(f'{txt} 标注了2个，且有GA，检测出大于2个 {classes} {classes2}') 63 | error[1] += 1 64 | print(f'[{i+1:>2}/{len(txts)}] {error} {out_path}/{txt}') 65 | 66 | return 0 67 | 68 | 69 | def get_P_from_label(self, path): 70 | txts = [x for x in os.listdir(path) if x.endswith('.txt')] 71 | assert len(txts) != 0 72 | """ 73 | 根据yolo的目标检测算法，每张图的样本总数应该是：7x7x3 74 | """ 75 | P = np.zeros(3) # crosswalk guide arrows total 76 | for i, txt in enumerate(txts): 77 | with open(f'{path}/{txt}', 'r') as f: 78 | data = f.readlines() 79 | classes = [x.split()[0] for x in data] # 标注文件中的类别 80 | assert len(classes) != 0 81 | for j, cls in enumerate(classes): 82 | if cls == '0': 83 | P[0] += 1 84 | elif cls == '1': 85 | P[1] += 1 86 | if len(classes) >=2: 87 | print(f'同时存在，{classes}') 88 | else: 89 | raise NameError(f'标注文件读取到的分类{cls}不正确，文件：{path}/{txt}') 90 | P[2] += (7*7*3 - len(classes)) 91 | 92 | print(f'\r[{i+1:>2}/{len(txts)}] {P} ', end='') 93 | 94 | P = P/np.sum(P) 95 | print(P) 96 | return P 97 | 98 | def cal_information_entropy(self, P): 99 | """ 100 | 输入的P是一个ndarray，长度是可能的符号数目，每个值是对应的概率Pxi 101 | H(x) = - sum(Pxi * log(Pxi)) 102 | :param P: 103 | :return: 104 | """ 105 | P = P[P != 0] # 筛选掉为0的项 106 | return -np.sum(np.multiply(P, np.log2(P))) 107 | 108 | def read_txt(self, txt): 109 | if os.path.exists(txt): 110 | with open(txt, 'r') as f: 111 | data = f.readlines() 112 | return [x.split()[0] for x in data] 113 | else: 114 | return None 115 | 116 | 117 | if __name__ == '__main__': 118 | ea = EntropyAnalyse() 119 | ea.run() 120 | -------------------------------------------------------------------------------- /for_paper/overall_data: -------------------------------------------------------------------------------- 1 | ts base btime base+NST base+NST+SE base+NST+ROI base+NST+SE+ROI base+NST+SE+ROI+SSVM 2 | 128 34.53 75 48.84 22.56 46 34.19 4.38 3 | 288 81.34 189 89.03 86.41 74 81.27 79.33 4 | 448 87.98 314 90.77 90.64 169 85.30 82.41 5 | 640 90.34 634 91.18 92.10 273 87.17 86.25 6 | 832 87.98 1134 90.66 90.17 403 88.85 87.69 7 | 8 | -------------------------------------------------------------------------------- /models/common.py: -------------------------------------------------------------------------------- 1 | # This file contains modules common to various models 2 | import math 3 | 4 | import torch 5 | import torch.nn as nn 6 | 7 | 8 | def autopad(k, p=None): # kernel, padding 9 | # Pad to 'same' 10 | if p is None: 11 | p = k // 2 if isinstance(k, int) else [x // 2 for x in k] # auto-pad 12 | return p 13 | 14 | 15 | def DWConv(c1, c2, k=1, s=1, act=True): 16 | # Depthwise convolution 17 | return Conv(c1, c2, k, s, g=math.gcd(c1, c2), act=act) 18 | 19 | 20 | class Conv(nn.Module): 21 | # Standard convolution 22 | def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups 23 | super(Conv, self).__init__() 24 | self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False) 25 | self.bn = nn.BatchNorm2d(c2) 26 | self.act = nn.LeakyReLU(0.1, inplace=True) if act else nn.Identity() 27 | 28 | def forward(self, x): 29 | return self.act(self.bn(self.conv(x))) 30 | 31 | def fuseforward(self, x): 32 | return self.act(self.conv(x)) 33 | 34 | 35 | class Bottleneck(nn.Module): 36 | # Standard bottleneck 37 | def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion 38 | super(Bottleneck, self).__init__() 39 | c_ = int(c2 * e) # hidden channels 40 | self.cv1 = Conv(c1, c_, 1, 1) 41 | self.cv2 = Conv(c_, c2, 3, 1, g=g) 42 | self.add = shortcut and c1 == c2 43 | 44 | def forward(self, x): 45 | return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) 46 | 47 | 48 | class BottleneckSE(nn.Module): 49 | """ 50 | 添加了SENet的bottleneck 51 | """ 52 | 53 | def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion 54 | super(Bottleneck, self).__init__() 55 | c_ = int(c2 * e) # hidden channels 56 | self.cv1 = Conv(c1, c_, 1, 1) 57 | # self.cv2 = Conv(c_, c2, 3, 1, g=g) 58 | 59 | self.conv = nn.Conv2d(c_, c2, 3, 1, autopad(3, None), groups=g, bias=False) 60 | self.bn = nn.BatchNorm2d(c2) 61 | 62 | self.se = SELayer(c2, reduction=16) 63 | self.act = nn.LeakyReLU(0.1, inplace=True) 64 | 65 | self.add = shortcut and c1 == c2 66 | 67 | def forward(self, x): 68 | residual = x 69 | x = self.cv1(x) 70 | x = self.conv(x) 71 | x = self.bn(x) 72 | if self.add: 73 | x = self.se(x) 74 | x = residual + x 75 | x = self.act(x) 76 | 77 | return x 78 | 79 | # return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) 80 | 81 | 82 | class SELayer(nn.Module): 83 | def __init__(self, channel, reduction=16): 84 | super(SELayer, self).__init__() 85 | self.avg_pool = nn.AdaptiveAvgPool2d(1) 86 | self.fc = nn.Sequential( 87 | nn.Linear(channel, channel // reduction, bias=False), 88 | nn.ReLU(inplace=True), 89 | nn.Linear(channel // reduction, channel, bias=False), 90 | nn.Sigmoid() 91 | ) 92 | 93 | def forward(self, x): 94 | b, c, _, _ = x.size() 95 | y = self.avg_pool(x).view(b, c) 96 | y = self.fc(y).view(b, c, 1, 1) 97 | return x * y.expand_as(x) 98 | 99 | 100 | class BottleneckCSP(nn.Module): 101 | # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks 102 | def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion 103 | super(BottleneckCSP, self).__init__() 104 | c_ = int(c2 * e) # hidden channels 105 | self.cv1 = Conv(c1, c_, 1, 1) 106 | self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False) 107 | self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False) 108 | self.cv4 = Conv(2 * c_, c2, 1, 1) 109 | self.bn = nn.BatchNorm2d(2 * c_) # applied to cat(cv2, cv3) 110 | self.act = nn.LeakyReLU(0.1, inplace=True) 111 | self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)]) 112 | 113 | def forward(self, x): 114 | y1 = self.cv3(self.m(self.cv1(x))) 115 | y2 = self.cv2(x) 116 | return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1)))) 117 | 118 | 119 | class SPP(nn.Module): 120 | # Spatial pyramid pooling layer used in YOLOv3-SPP 121 | def __init__(self, c1, c2, k=(5, 9, 13)): 122 | super(SPP, self).__init__() 123 | c_ = c1 // 2 # hidden channels 124 | self.cv1 = Conv(c1, c_, 1, 1) 125 | self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1) 126 | self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k]) 127 | 128 | def forward(self, x): 129 | x = self.cv1(x) 130 | return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1)) 131 | 132 | 133 | class Focus(nn.Module): 134 | # Focus wh information into c-space 135 | def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups 136 | super(Focus, self).__init__() 137 | self.conv = Conv(c1 * 4, c2, k, s, p, g, act) 138 | 139 | def forward(self, x): # x(b,c,w,h) -> y(b,4c,w/2,h/2) 140 | return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1)) 141 | 142 | 143 | class Concat(nn.Module): 144 | # Concatenate a list of tensors along dimension 145 | def __init__(self, dimension=1): 146 | super(Concat, self).__init__() 147 | self.d = dimension 148 | 149 | def forward(self, x): 150 | return torch.cat(x, self.d) 151 | 152 | 153 | class Flatten(nn.Module): 154 | # Use after nn.AdaptiveAvgPool2d(1) to remove last 2 dimensions 155 | @staticmethod 156 | def forward(x): 157 | return x.view(x.size(0), -1) 158 | 159 | 160 | class Classify(nn.Module): 161 | # Classification head, i.e. x(b,c1,20,20) to x(b,c2) 162 | def __init__(self, c1, c2, k=1, s=1, p=None, g=1): # ch_in, ch_out, kernel, stride, padding, groups 163 | super(Classify, self).__init__() 164 | self.aap = nn.AdaptiveAvgPool2d(1) # to x(b,c1,1,1) 165 | self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False) # to x(b,c2,1,1) 166 | self.flat = Flatten() 167 | 168 | def forward(self, x): 169 | z = torch.cat([self.aap(y) for y in (x if isinstance(x, list) else [x])], 1) # cat if list 170 | return self.flat(self.conv(z)) # flatten to x(b,c2) 171 | -------------------------------------------------------------------------------- /models/commonSE.py: -------------------------------------------------------------------------------- 1 | # This file contains modules common to various models 2 | import math 3 | 4 | import torch 5 | import torch.nn as nn 6 | 7 | 8 | def autopad(k, p=None): # kernel, padding 9 | # Pad to 'same' 10 | if p is None: 11 | p = k // 2 if isinstance(k, int) else [x // 2 for x in k] # auto-pad 12 | return p 13 | 14 | 15 | def DWConv(c1, c2, k=1, s=1, act=True): 16 | # Depthwise convolution 17 | return Conv(c1, c2, k, s, g=math.gcd(c1, c2), act=act) 18 | 19 | 20 | class Conv(nn.Module): 21 | # Standard convolution 22 | def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups 23 | super(Conv, self).__init__() 24 | self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False) 25 | self.bn = nn.BatchNorm2d(c2) 26 | self.act = nn.LeakyReLU(0.1, inplace=True) if act else nn.Identity() 27 | 28 | def forward(self, x): 29 | return self.act(self.bn(self.conv(x))) 30 | 31 | def fuseforward(self, x): 32 | return self.act(self.conv(x)) 33 | 34 | 35 | class Bottleneck_back(nn.Module): 36 | # Standard bottleneck 37 | def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion 38 | super(Bottleneck, self).__init__() 39 | c_ = int(c2 * e) # hidden channels 40 | self.cv1 = Conv(c1, c_, 1, 1) 41 | self.cv2 = Conv(c_, c2, 3, 1, g=g) 42 | self.add = shortcut and c1 == c2 43 | 44 | def forward(self, x): 45 | return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) 46 | 47 | 48 | class Bottleneck(nn.Module): 49 | """ 50 | 添加了SENet的bottleneck 51 | """ 52 | 53 | def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion 54 | super(Bottleneck, self).__init__() 55 | c_ = int(c2 * e) # hidden channels 56 | self.cv1 = Conv(c1, c_, 1, 1) 57 | # self.cv2 = Conv(c_, c2, 3, 1, g=g) 58 | 59 | self.conv = nn.Conv2d(c_, c2, 3, 1, autopad(3, None), groups=g, bias=False) 60 | self.bn = nn.BatchNorm2d(c2) 61 | 62 | self.se = SELayer(c2, reduction=16) 63 | self.act = nn.LeakyReLU(0.1, inplace=True) 64 | 65 | self.add = shortcut and c1 == c2 66 | 67 | def forward(self, x): 68 | residual = x 69 | x = self.cv1(x) 70 | x = self.conv(x) 71 | x = self.bn(x) 72 | if self.add: 73 | x = self.se(x) 74 | x = residual + x 75 | x = self.act(x) 76 | 77 | return x 78 | # return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) 79 | 80 | 81 | class SELayer(nn.Module): 82 | def __init__(self, channel, reduction=16): 83 | super(SELayer, self).__init__() 84 | self.avg_pool = nn.AdaptiveAvgPool2d(1) 85 | self.fc = nn.Sequential( 86 | nn.Linear(channel, channel // reduction, bias=False), 87 | nn.ReLU(inplace=True), 88 | nn.Linear(channel // reduction, channel, bias=False), 89 | nn.Sigmoid() 90 | ) 91 | 92 | def forward(self, x): 93 | b, c, _, _ = x.size() 94 | y = self.avg_pool(x).view(b, c) 95 | y = self.fc(y).view(b, c, 1, 1) 96 | return x * y.expand_as(x) 97 | 98 | 99 | class BottleneckCSP(nn.Module): 100 | # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks 101 | def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion 102 | super(BottleneckCSP, self).__init__() 103 | c_ = int(c2 * e) # hidden channels 104 | self.cv1 = Conv(c1, c_, 1, 1) 105 | self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False) 106 | self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False) 107 | self.cv4 = Conv(2 * c_, c2, 1, 1) 108 | self.bn = nn.BatchNorm2d(2 * c_) # applied to cat(cv2, cv3) 109 | self.act = nn.LeakyReLU(0.1, inplace=True) 110 | self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)]) 111 | 112 | def forward(self, x): 113 | y1 = self.cv3(self.m(self.cv1(x))) 114 | y2 = self.cv2(x) 115 | return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1)))) 116 | 117 | 118 | class SPP(nn.Module): 119 | # Spatial pyramid pooling layer used in YOLOv3-SPP 120 | def __init__(self, c1, c2, k=(5, 9, 13)): 121 | super(SPP, self).__init__() 122 | c_ = c1 // 2 # hidden channels 123 | self.cv1 = Conv(c1, c_, 1, 1) 124 | self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1) 125 | self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k]) 126 | 127 | def forward(self, x): 128 | x = self.cv1(x) 129 | return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1)) 130 | 131 | 132 | class Focus(nn.Module): 133 | # Focus wh information into c-space 134 | def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups 135 | super(Focus, self).__init__() 136 | self.conv = Conv(c1 * 4, c2, k, s, p, g, act) 137 | 138 | def forward(self, x): # x(b,c,w,h) -> y(b,4c,w/2,h/2) 139 | return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1)) 140 | 141 | 142 | class Concat(nn.Module): 143 | # Concatenate a list of tensors along dimension 144 | def __init__(self, dimension=1): 145 | super(Concat, self).__init__() 146 | self.d = dimension 147 | 148 | def forward(self, x): 149 | return torch.cat(x, self.d) 150 | 151 | 152 | class Flatten(nn.Module): 153 | # Use after nn.AdaptiveAvgPool2d(1) to remove last 2 dimensions 154 | @staticmethod 155 | def forward(x): 156 | return x.view(x.size(0), -1) 157 | 158 | 159 | class Classify(nn.Module): 160 | # Classification head, i.e. x(b,c1,20,20) to x(b,c2) 161 | def __init__(self, c1, c2, k=1, s=1, p=None, g=1): # ch_in, ch_out, kernel, stride, padding, groups 162 | super(Classify, self).__init__() 163 | self.aap = nn.AdaptiveAvgPool2d(1) # to x(b,c1,1,1) 164 | self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False) # to x(b,c2,1,1) 165 | self.flat = Flatten() 166 | 167 | def forward(self, x): 168 | z = torch.cat([self.aap(y) for y in (x if isinstance(x, list) else [x])], 1) # cat if list 169 | return self.flat(self.conv(z)) # flatten to x(b,c2) 170 | -------------------------------------------------------------------------------- /models/experimental.py: -------------------------------------------------------------------------------- 1 | # This file contains experimental modules 2 | 3 | import numpy as np 4 | import torch 5 | import torch.nn as nn 6 | 7 | from models.common import Conv, DWConv 8 | from utils.google_utils import attempt_download 9 | 10 | 11 | class CrossConv(nn.Module): 12 | # Cross Convolution Downsample 13 | def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False): 14 | # ch_in, ch_out, kernel, stride, groups, expansion, shortcut 15 | super(CrossConv, self).__init__() 16 | c_ = int(c2 * e) # hidden channels 17 | self.cv1 = Conv(c1, c_, (1, k), (1, s)) 18 | self.cv2 = Conv(c_, c2, (k, 1), (s, 1), g=g) 19 | self.add = shortcut and c1 == c2 20 | 21 | def forward(self, x): 22 | return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) 23 | 24 | 25 | class C3(nn.Module): 26 | # Cross Convolution CSP 27 | def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion 28 | super(C3, self).__init__() 29 | c_ = int(c2 * e) # hidden channels 30 | self.cv1 = Conv(c1, c_, 1, 1) 31 | self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False) 32 | self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False) 33 | self.cv4 = Conv(2 * c_, c2, 1, 1) 34 | self.bn = nn.BatchNorm2d(2 * c_) # applied to cat(cv2, cv3) 35 | self.act = nn.LeakyReLU(0.1, inplace=True) 36 | self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)]) 37 | 38 | def forward(self, x): 39 | y1 = self.cv3(self.m(self.cv1(x))) 40 | y2 = self.cv2(x) 41 | return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1)))) 42 | 43 | 44 | class Sum(nn.Module): 45 | # Weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070 46 | def __init__(self, n, weight=False): # n: number of inputs 47 | super(Sum, self).__init__() 48 | self.weight = weight # apply weights boolean 49 | self.iter = range(n - 1) # iter object 50 | if weight: 51 | self.w = nn.Parameter(-torch.arange(1., n) / 2, requires_grad=True) # layer weights 52 | 53 | def forward(self, x): 54 | y = x[0] # no weight 55 | if self.weight: 56 | w = torch.sigmoid(self.w) * 2 57 | for i in self.iter: 58 | y = y + x[i + 1] * w[i] 59 | else: 60 | for i in self.iter: 61 | y = y + x[i + 1] 62 | return y 63 | 64 | 65 | class GhostConv(nn.Module): 66 | # Ghost Convolution https://github.com/huawei-noah/ghostnet 67 | def __init__(self, c1, c2, k=1, s=1, g=1, act=True): # ch_in, ch_out, kernel, stride, groups 68 | super(GhostConv, self).__init__() 69 | c_ = c2 // 2 # hidden channels 70 | self.cv1 = Conv(c1, c_, k, s, g, act) 71 | self.cv2 = Conv(c_, c_, 5, 1, c_, act) 72 | 73 | def forward(self, x): 74 | y = self.cv1(x) 75 | return torch.cat([y, self.cv2(y)], 1) 76 | 77 | 78 | class GhostBottleneck(nn.Module): 79 | # Ghost Bottleneck https://github.com/huawei-noah/ghostnet 80 | def __init__(self, c1, c2, k, s): 81 | super(GhostBottleneck, self).__init__() 82 | c_ = c2 // 2 83 | self.conv = nn.Sequential(GhostConv(c1, c_, 1, 1), # pw 84 | DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(), # dw 85 | GhostConv(c_, c2, 1, 1, act=False)) # pw-linear 86 | self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False), 87 | Conv(c1, c2, 1, 1, act=False)) if s == 2 else nn.Identity() 88 | 89 | def forward(self, x): 90 | return self.conv(x) + self.shortcut(x) 91 | 92 | 93 | class MixConv2d(nn.Module): 94 | # Mixed Depthwise Conv https://arxiv.org/abs/1907.09595 95 | def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True): 96 | super(MixConv2d, self).__init__() 97 | groups = len(k) 98 | if equal_ch: # equal c_ per group 99 | i = torch.linspace(0, groups - 1E-6, c2).floor() # c2 indices 100 | c_ = [(i == g).sum() for g in range(groups)] # intermediate channels 101 | else: # equal weight.numel() per group 102 | b = [c2] + [0] * groups 103 | a = np.eye(groups + 1, groups, k=-1) 104 | a -= np.roll(a, 1, axis=1) 105 | a *= np.array(k) ** 2 106 | a[0] = 1 107 | c_ = np.linalg.lstsq(a, b, rcond=None)[0].round() # solve for equal weight indices, ax = b 108 | 109 | self.m = nn.ModuleList([nn.Conv2d(c1, int(c_[g]), k[g], s, k[g] // 2, bias=False) for g in range(groups)]) 110 | self.bn = nn.BatchNorm2d(c2) 111 | self.act = nn.LeakyReLU(0.1, inplace=True) 112 | 113 | def forward(self, x): 114 | return x + self.act(self.bn(torch.cat([m(x) for m in self.m], 1))) 115 | 116 | 117 | class Ensemble(nn.ModuleList): 118 | # Ensemble of models 119 | def __init__(self): 120 | super(Ensemble, self).__init__() 121 | 122 | def forward(self, x, augment=False): 123 | y = [] 124 | for module in self: 125 | y.append(module(x, augment)[0]) 126 | # y = torch.stack(y).max(0)[0] # max ensemble 127 | # y = torch.cat(y, 1) # nms ensemble 128 | y = torch.stack(y).mean(0) # mean ensemble 129 | return y, None # inference, train output 130 | 131 | 132 | def attempt_load(weights, map_location=None): 133 | # Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a 134 | model = Ensemble() 135 | for w in weights if isinstance(weights, list) else [weights]: 136 | attempt_download(w) 137 | model.append(torch.load(w, map_location=map_location)['model'].float().fuse().eval()) # load FP32 model 138 | 139 | if len(model) == 1: 140 | return model[-1] # return model 141 | else: 142 | print('Ensemble created with %s\n' % weights) 143 | for k in ['names', 'stride']: 144 | setattr(model, k, getattr(model[-1], k)) 145 | return model # return ensemble 146 | -------------------------------------------------------------------------------- /models/yolov5m.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 2 # number of classes 3 | depth_multiple: 0.67 # model depth multiple 4 | width_multiple: 0.75 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, BottleneckCSP, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, BottleneckCSP, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, BottleneckCSP, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, BottleneckCSP, [1024, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 32 | [-1, 3, BottleneckCSP, [512, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 37 | [-1, 3, BottleneckCSP, [256, False]], # 17 (P3/8-small) 38 | 39 | [-1, 1, Conv, [256, 3, 2]], 40 | [[-1, 14], 1, Concat, [1]], # cat head P4 41 | [-1, 3, BottleneckCSP, [512, False]], # 20 (P4/16-medium) 42 | 43 | [-1, 1, Conv, [512, 3, 2]], 44 | [[-1, 10], 1, Concat, [1]], # cat head P5 45 | [-1, 3, BottleneckCSP, [1024, False]], # 23 (P5/32-large) 46 | 47 | [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 48 | ] 49 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/_base_/datasets/zzd_coco_detection.py: -------------------------------------------------------------------------------- 1 | # dataset settings 2 | dataset_type = 'CocoDataset' 3 | data_root = '/home/zzd/datasets/mscoco/' 4 | img_norm_cfg = dict( 5 | mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) 6 | train_pipeline = [ 7 | dict(type='LoadImageFromFile'), 8 | dict(type='LoadAnnotations', with_bbox=True), 9 | dict(type='Resize', img_scale=(1333, 800), keep_ratio=True), 10 | dict(type='RandomFlip', flip_ratio=0.5), 11 | dict(type='Normalize', **img_norm_cfg), 12 | dict(type='Pad', size_divisor=32), 13 | dict(type='DefaultFormatBundle'), 14 | dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']), 15 | ] 16 | test_pipeline = [ 17 | dict(type='LoadImageFromFile'), 18 | dict( 19 | type='MultiScaleFlipAug', 20 | img_scale=(1333, 800), 21 | flip=False, 22 | transforms=[ 23 | dict(type='Resize', keep_ratio=True), 24 | dict(type='RandomFlip'), 25 | dict(type='Normalize', **img_norm_cfg), 26 | dict(type='Pad', size_divisor=32), 27 | dict(type='ImageToTensor', keys=['img']), 28 | dict(type='Collect', keys=['img']), 29 | ]) 30 | ] 31 | data = dict( 32 | samples_per_gpu=2, 33 | workers_per_gpu=2, 34 | train=dict( 35 | type=dataset_type, 36 | ann_file=data_root + 'annotations/instances_train2017.json', 37 | img_prefix=data_root + 'train2017/', 38 | pipeline=train_pipeline), 39 | val=dict( 40 | type=dataset_type, 41 | ann_file=data_root + 'annotations/instances_val2017.json', 42 | img_prefix=data_root + 'val2017/', 43 | pipeline=test_pipeline), 44 | test=dict( 45 | type=dataset_type, 46 | ann_file=data_root + 'annotations/instances_val2017.json', 47 | img_prefix=data_root + 'val2017/', 48 | pipeline=test_pipeline)) 49 | evaluation = dict(interval=1, metric='bbox') 50 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/_base_/datasets/zzd_coco_instance.py: -------------------------------------------------------------------------------- 1 | # dataset settings 2 | dataset_type = 'CocoDataset' 3 | data_root = '/home/zzd/datasets/mscoco/' 4 | img_norm_cfg = dict( 5 | mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) 6 | train_pipeline = [ 7 | dict(type='LoadImageFromFile'), 8 | dict(type='LoadAnnotations', with_bbox=True, with_mask=True), 9 | dict(type='Resize', img_scale=(1333, 800), keep_ratio=True), 10 | dict(type='RandomFlip', flip_ratio=0.5), 11 | dict(type='Normalize', **img_norm_cfg), 12 | dict(type='Pad', size_divisor=32), 13 | dict(type='DefaultFormatBundle'), 14 | dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks']), 15 | ] 16 | test_pipeline = [ 17 | dict(type='LoadImageFromFile'), 18 | dict( 19 | type='MultiScaleFlipAug', 20 | img_scale=(1333, 800), 21 | flip=False, 22 | transforms=[ 23 | dict(type='Resize', keep_ratio=True), 24 | dict(type='RandomFlip'), 25 | dict(type='Normalize', **img_norm_cfg), 26 | dict(type='Pad', size_divisor=32), 27 | dict(type='ImageToTensor', keys=['img']), 28 | dict(type='Collect', keys=['img']), 29 | ]) 30 | ] 31 | data = dict( 32 | samples_per_gpu=2, 33 | workers_per_gpu=2, 34 | train=dict( 35 | type=dataset_type, 36 | ann_file=data_root + 'annotations/instances_train2017.json', 37 | img_prefix=data_root + 'train2017/', 38 | pipeline=train_pipeline), 39 | val=dict( 40 | type=dataset_type, 41 | ann_file=data_root + 'annotations/instances_val2017.json', 42 | img_prefix=data_root + 'val2017/', 43 | pipeline=test_pipeline), 44 | test=dict( 45 | type=dataset_type, 46 | ann_file=data_root + 'annotations/instances_val2017.json', 47 | img_prefix=data_root + 'val2017/', 48 | pipeline=test_pipeline)) 49 | evaluation = dict(metric=['bbox', 'segm']) 50 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/_base_/datasets/zzd_crosswalk_detection.py: -------------------------------------------------------------------------------- 1 | # dataset settings 2 | dataset_type = 'CocoDataset' 3 | classes = ('crosswalk', 'guide_arrows') 4 | data_root = '/home/zzd/datasets/crosswalk/fogged_train_data_coco_format/' 5 | img_norm_cfg = dict( 6 | mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) 7 | train_pipeline = [ 8 | dict(type='LoadImageFromFile'), 9 | dict(type='LoadAnnotations', with_bbox=True), 10 | dict(type='Resize', img_scale=(1333, 800), keep_ratio=True), 11 | dict(type='RandomFlip', flip_ratio=0.5), 12 | dict(type='Normalize', **img_norm_cfg), 13 | dict(type='Pad', size_divisor=32), 14 | dict(type='DefaultFormatBundle'), 15 | dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']), 16 | ] 17 | test_pipeline = [ 18 | dict(type='LoadImageFromFile'), 19 | dict( 20 | type='MultiScaleFlipAug', 21 | img_scale=(1333, 800), 22 | flip=False, 23 | transforms=[ 24 | dict(type='Resize', keep_ratio=True), 25 | dict(type='RandomFlip'), 26 | dict(type='Normalize', **img_norm_cfg), 27 | dict(type='Pad', size_divisor=32), 28 | dict(type='ImageToTensor', keys=['img']), 29 | dict(type='Collect', keys=['img']), 30 | ]) 31 | ] 32 | data = dict( 33 | samples_per_gpu=2, 34 | workers_per_gpu=2, 35 | train=dict( 36 | type=dataset_type, 37 | ann_file=data_root + 'annotations/instances_train2017.json', 38 | img_prefix=data_root + 'train2017/', 39 | classes=classes, 40 | pipeline=train_pipeline), 41 | val=dict( 42 | type=dataset_type, 43 | ann_file=data_root + 'annotations/instances_val2017.json', 44 | img_prefix=data_root + 'val2017/', 45 | classes=classes, 46 | pipeline=test_pipeline), 47 | test=dict( 48 | type=dataset_type, 49 | ann_file=data_root + 'annotations/instances_val2017.json', 50 | img_prefix=data_root + 'val2017/', 51 | classes=classes, 52 | pipeline=test_pipeline)) 53 | evaluation = dict(interval=1, metric='bbox') 54 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/_base_/datasets/zzd_crosswalk_instance.py: -------------------------------------------------------------------------------- 1 | # dataset settings 2 | dataset_type = 'CocoDataset' 3 | classes = ('crosswalk', 'guide_arrows') 4 | data_root = '/home/zzd/datasets/crosswalk/fogged_train_data_coco_format/' 5 | img_norm_cfg = dict( 6 | mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) 7 | train_pipeline = [ 8 | dict(type='LoadImageFromFile'), 9 | dict(type='LoadAnnotations', with_bbox=True, with_mask=True), 10 | dict(type='Resize', img_scale=(1333, 800), keep_ratio=True), 11 | dict(type='RandomFlip', flip_ratio=0.5), 12 | dict(type='Normalize', **img_norm_cfg), 13 | dict(type='Pad', size_divisor=32), 14 | dict(type='DefaultFormatBundle'), 15 | dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks']), 16 | ] 17 | test_pipeline = [ 18 | dict(type='LoadImageFromFile'), 19 | dict( 20 | type='MultiScaleFlipAug', 21 | img_scale=(1333, 800), 22 | flip=False, 23 | transforms=[ 24 | dict(type='Resize', keep_ratio=True), 25 | dict(type='RandomFlip'), 26 | dict(type='Normalize', **img_norm_cfg), 27 | dict(type='Pad', size_divisor=32), 28 | dict(type='ImageToTensor', keys=['img']), 29 | dict(type='Collect', keys=['img']), 30 | ]) 31 | ] 32 | data = dict( 33 | samples_per_gpu=2, 34 | workers_per_gpu=2, 35 | train=dict( 36 | type=dataset_type, 37 | ann_file=data_root + 'annotations/instances_train2017.json', 38 | img_prefix=data_root + 'train2017/', 39 | classes=classes, 40 | pipeline=train_pipeline), 41 | val=dict( 42 | type=dataset_type, 43 | ann_file=data_root + 'annotations/instances_val2017.json', 44 | img_prefix=data_root + 'val2017/', 45 | classes=classes, 46 | pipeline=test_pipeline), 47 | test=dict( 48 | type=dataset_type, 49 | ann_file=data_root + 'annotations/instances_val2017.json', 50 | img_prefix=data_root + 'val2017/', 51 | classes=classes, 52 | pipeline=test_pipeline)) 53 | evaluation = dict(metric=['bbox', 'segm']) 54 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/centernet/zzd_centernet_resnet18_dcnv2_140e_crosswalk.py: -------------------------------------------------------------------------------- 1 | _base_ = [ 2 | '../_base_/datasets/zzd_crosswalk_detection.py', 3 | '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py' 4 | ] 5 | 6 | model = dict( 7 | type='CenterNet', 8 | backbone=dict( 9 | type='ResNet', 10 | depth=18, 11 | norm_eval=False, 12 | norm_cfg=dict(type='BN'), 13 | init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet18')), 14 | neck=dict( 15 | type='CTResNetNeck', 16 | in_channel=512, 17 | num_deconv_filters=(256, 128, 64), 18 | num_deconv_kernels=(4, 4, 4), 19 | use_dcn=True), 20 | bbox_head=dict( 21 | type='CenterNetHead', 22 | num_classes=2, 23 | in_channel=64, 24 | feat_channel=64, 25 | loss_center_heatmap=dict(type='GaussianFocalLoss', loss_weight=1.0), 26 | loss_wh=dict(type='L1Loss', loss_weight=0.1), 27 | loss_offset=dict(type='L1Loss', loss_weight=1.0)), 28 | train_cfg=None, 29 | test_cfg=dict(topk=100, local_maximum_kernel=3, max_per_img=100)) 30 | 31 | # We fixed the incorrect img_norm_cfg problem in the source code. 32 | img_norm_cfg = dict( 33 | mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) 34 | 35 | train_pipeline = [ 36 | dict(type='LoadImageFromFile', to_float32=True, color_type='color'), 37 | dict(type='LoadAnnotations', with_bbox=True), 38 | dict( 39 | type='PhotoMetricDistortion', 40 | brightness_delta=32, 41 | contrast_range=(0.5, 1.5), 42 | saturation_range=(0.5, 1.5), 43 | hue_delta=18), 44 | dict( 45 | type='RandomCenterCropPad', 46 | crop_size=(512, 512), 47 | ratios=(0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3), 48 | mean=[0, 0, 0], 49 | std=[1, 1, 1], 50 | to_rgb=True, 51 | test_pad_mode=None), 52 | dict(type='Resize', img_scale=(512, 512), keep_ratio=True), 53 | dict(type='RandomFlip', flip_ratio=0.5), 54 | dict(type='Normalize', **img_norm_cfg), 55 | dict(type='DefaultFormatBundle'), 56 | dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']) 57 | ] 58 | test_pipeline = [ 59 | dict(type='LoadImageFromFile', to_float32=True), 60 | dict( 61 | type='MultiScaleFlipAug', 62 | scale_factor=1.0, 63 | flip=False, 64 | transforms=[ 65 | dict(type='Resize', keep_ratio=True), 66 | dict( 67 | type='RandomCenterCropPad', 68 | ratios=None, 69 | border=None, 70 | mean=[0, 0, 0], 71 | std=[1, 1, 1], 72 | to_rgb=True, 73 | test_mode=True, 74 | test_pad_mode=['logical_or', 31], 75 | test_pad_add_pix=1), 76 | dict(type='RandomFlip'), 77 | dict(type='Normalize', **img_norm_cfg), 78 | dict(type='DefaultFormatBundle'), 79 | dict( 80 | type='Collect', 81 | meta_keys=('filename', 'ori_shape', 'img_shape', 'pad_shape', 82 | 'scale_factor', 'flip', 'flip_direction', 83 | 'img_norm_cfg', 'border'), 84 | keys=['img']) 85 | ]) 86 | ] 87 | 88 | classes = ('crosswalk', 'guide_arrows') 89 | dataset_type = 'CocoDataset' 90 | # data_root = 'data/coco/' 91 | data_root = '/home/zzd/datasets/crosswalk/fogged_train_data_coco_format/' 92 | 93 | # Use RepeatDataset to speed up training 94 | data = dict( 95 | samples_per_gpu=8, 96 | workers_per_gpu=2, 97 | train=dict( 98 | _delete_=True, 99 | type='RepeatDataset', 100 | times=5, 101 | dataset=dict( 102 | type=dataset_type, 103 | ann_file=data_root + 'annotations/instances_train2017.json', 104 | img_prefix=data_root + 'train2017/', 105 | pipeline=train_pipeline, classes=classes)), 106 | val=dict(pipeline=test_pipeline, classes=classes), 107 | test=dict(pipeline=test_pipeline, classes=classes)) 108 | 109 | # optimizer 110 | # Based on the default settings of modern detectors, the SGD effect is better 111 | # than the Adam in the source code, so we use SGD default settings and 112 | # if you use adam+lr5e-4, the map is 29.1. 113 | optimizer_config = dict( 114 | _delete_=True, grad_clip=dict(max_norm=35, norm_type=2)) 115 | 116 | # learning policy 117 | # Based on the default settings of modern detectors, we added warmup settings. 118 | lr_config = dict( 119 | policy='step', 120 | warmup='linear', 121 | warmup_iters=1000, 122 | warmup_ratio=1.0 / 1000, 123 | step=[18, 24]) # the real step is [18*5, 24*5] 124 | runner = dict(max_epochs=28) # the real epoch is 28*5=140 125 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/deformable_detr/zzd_deformable_detr_r50_16x2_50e_crosswalk.py: -------------------------------------------------------------------------------- 1 | _base_ = [ 2 | '../_base_/datasets/zzd_crosswalk_detection.py', '../_base_/default_runtime.py' 3 | ] 4 | model = dict( 5 | type='DeformableDETR', 6 | backbone=dict( 7 | type='ResNet', 8 | depth=50, 9 | num_stages=4, 10 | out_indices=(1, 2, 3), 11 | frozen_stages=1, 12 | norm_cfg=dict(type='BN', requires_grad=False), 13 | norm_eval=True, 14 | style='pytorch', 15 | init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50')), 16 | neck=dict( 17 | type='ChannelMapper', 18 | in_channels=[512, 1024, 2048], 19 | kernel_size=1, 20 | out_channels=256, 21 | act_cfg=None, 22 | norm_cfg=dict(type='GN', num_groups=32), 23 | num_outs=4), 24 | bbox_head=dict( 25 | type='DeformableDETRHead', 26 | num_query=300, 27 | num_classes=2, 28 | in_channels=2048, 29 | sync_cls_avg_factor=True, 30 | as_two_stage=False, 31 | transformer=dict( 32 | type='DeformableDetrTransformer', 33 | encoder=dict( 34 | type='DetrTransformerEncoder', 35 | num_layers=6, 36 | transformerlayers=dict( 37 | type='BaseTransformerLayer', 38 | attn_cfgs=dict( 39 | type='MultiScaleDeformableAttention', embed_dims=256), 40 | feedforward_channels=1024, 41 | ffn_dropout=0.1, 42 | operation_order=('self_attn', 'norm', 'ffn', 'norm'))), 43 | decoder=dict( 44 | type='DeformableDetrTransformerDecoder', 45 | num_layers=6, 46 | return_intermediate=True, 47 | transformerlayers=dict( 48 | type='DetrTransformerDecoderLayer', 49 | attn_cfgs=[ 50 | dict( 51 | type='MultiheadAttention', 52 | embed_dims=256, 53 | num_heads=8, 54 | dropout=0.1), 55 | dict( 56 | type='MultiScaleDeformableAttention', 57 | embed_dims=256) 58 | ], 59 | feedforward_channels=1024, 60 | ffn_dropout=0.1, 61 | operation_order=('self_attn', 'norm', 'cross_attn', 'norm', 62 | 'ffn', 'norm')))), 63 | positional_encoding=dict( 64 | type='SinePositionalEncoding', 65 | num_feats=128, 66 | normalize=True, 67 | offset=-0.5), 68 | loss_cls=dict( 69 | type='FocalLoss', 70 | use_sigmoid=True, 71 | gamma=2.0, 72 | alpha=0.25, 73 | loss_weight=2.0), 74 | loss_bbox=dict(type='L1Loss', loss_weight=5.0), 75 | loss_iou=dict(type='GIoULoss', loss_weight=2.0)), 76 | # training and testing settings 77 | train_cfg=dict( 78 | assigner=dict( 79 | type='HungarianAssigner', 80 | cls_cost=dict(type='FocalLossCost', weight=2.0), 81 | reg_cost=dict(type='BBoxL1Cost', weight=5.0, box_format='xywh'), 82 | iou_cost=dict(type='IoUCost', iou_mode='giou', weight=2.0))), 83 | test_cfg=dict(max_per_img=100)) 84 | img_norm_cfg = dict( 85 | mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) 86 | # train_pipeline, NOTE the img_scale and the Pad's size_divisor is different 87 | # from the default setting in mmdet. 88 | train_pipeline = [ 89 | dict(type='LoadImageFromFile'), 90 | dict(type='LoadAnnotations', with_bbox=True), 91 | dict(type='RandomFlip', flip_ratio=0.5), 92 | dict( 93 | type='AutoAugment', 94 | policies=[ 95 | [ 96 | dict( 97 | type='Resize', 98 | img_scale=[(480, 1333), (512, 1333), (544, 1333), 99 | (576, 1333), (608, 1333), (640, 1333), 100 | (672, 1333), (704, 1333), (736, 1333), 101 | (768, 1333), (800, 1333)], 102 | multiscale_mode='value', 103 | keep_ratio=True) 104 | ], 105 | [ 106 | dict( 107 | type='Resize', 108 | # The radio of all image in train dataset < 7 109 | # follow the original impl 110 | img_scale=[(400, 4200), (500, 4200), (600, 4200)], 111 | multiscale_mode='value', 112 | keep_ratio=True), 113 | dict( 114 | type='RandomCrop', 115 | crop_type='absolute_range', 116 | crop_size=(384, 600), 117 | allow_negative_crop=True), 118 | dict( 119 | type='Resize', 120 | img_scale=[(480, 1333), (512, 1333), (544, 1333), 121 | (576, 1333), (608, 1333), (640, 1333), 122 | (672, 1333), (704, 1333), (736, 1333), 123 | (768, 1333), (800, 1333)], 124 | multiscale_mode='value', 125 | override=True, 126 | keep_ratio=True) 127 | ] 128 | ]), 129 | dict(type='Normalize', **img_norm_cfg), 130 | dict(type='Pad', size_divisor=1), 131 | dict(type='DefaultFormatBundle'), 132 | dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']) 133 | ] 134 | # test_pipeline, NOTE the Pad's size_divisor is different from the default 135 | # setting (size_divisor=32). While there is little effect on the performance 136 | # whether we use the default setting or use size_divisor=1. 137 | test_pipeline = [ 138 | dict(type='LoadImageFromFile'), 139 | dict( 140 | type='MultiScaleFlipAug', 141 | img_scale=(1333, 800), 142 | flip=False, 143 | transforms=[ 144 | dict(type='Resize', keep_ratio=True), 145 | dict(type='RandomFlip'), 146 | dict(type='Normalize', **img_norm_cfg), 147 | dict(type='Pad', size_divisor=1), 148 | dict(type='ImageToTensor', keys=['img']), 149 | dict(type='Collect', keys=['img']) 150 | ]) 151 | ] 152 | 153 | classes = ('crosswalk', 'guide_arrows') 154 | data = dict( 155 | samples_per_gpu=1, 156 | workers_per_gpu=2, 157 | train=dict(filter_empty_gt=False, pipeline=train_pipeline, classes=classes), 158 | val=dict(pipeline=test_pipeline, classes=classes), 159 | test=dict(pipeline=test_pipeline, classes=classes)) 160 | # optimizer 161 | optimizer = dict( 162 | type='Adam', 163 | lr=2e-4, 164 | weight_decay=0.0001, 165 | paramwise_cfg=dict( 166 | custom_keys={ 167 | 'backbone': dict(lr_mult=0.1), 168 | 'sampling_offsets': dict(lr_mult=0.1), 169 | 'reference_points': dict(lr_mult=0.1) 170 | })) 171 | optimizer_config = dict(grad_clip=dict(max_norm=0.1, norm_type=2)) 172 | # learning policy 173 | lr_config = dict(policy='step', step=[40]) 174 | runner = dict(type='EpochBasedRunner', max_epochs=50) 175 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/faster_rcnn/zzd_faster_rcnn_r50_fpn_1x_coco.py: -------------------------------------------------------------------------------- 1 | _base_ = [ 2 | '../_base_/models/faster_rcnn_r50_fpn.py', 3 | '../_base_/datasets/zzd_crosswalk_detection.py', 4 | '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py' 5 | ] 6 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/faster_rcnn/zzd_faster_rcnn_r50_fpn_1x_crosswalk.py: -------------------------------------------------------------------------------- 1 | _base_ = [ 2 | '../_base_/models/faster_rcnn_r50_fpn.py', 3 | '../_base_/datasets/zzd_crosswalk_detection.py', 4 | '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py' 5 | ] 6 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/htc/zzd_htc_r50_fpn_1x_crosswalk.py: -------------------------------------------------------------------------------- 1 | _base_ = './zzd_htc_without_semantic_r50_fpn_1x_crosswalk.py' 2 | model = dict( 3 | roi_head=dict( 4 | semantic_roi_extractor=dict( 5 | type='SingleRoIExtractor', 6 | roi_layer=dict(type='RoIAlign', output_size=14, sampling_ratio=0), 7 | out_channels=256, 8 | featmap_strides=[8]), 9 | semantic_head=dict( 10 | type='FusedSemanticHead', 11 | num_ins=5, 12 | fusion_level=1, 13 | num_convs=4, 14 | in_channels=256, 15 | conv_out_channels=256, 16 | num_classes=183, 17 | loss_seg=dict( 18 | type='CrossEntropyLoss', ignore_index=255, loss_weight=0.2)))) 19 | data_root = 'data/coco/' 20 | img_norm_cfg = dict( 21 | mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) 22 | train_pipeline = [ 23 | dict(type='LoadImageFromFile'), 24 | dict( 25 | type='LoadAnnotations', with_bbox=True, with_mask=True, with_seg=True), 26 | dict(type='Resize', img_scale=(1333, 800), keep_ratio=True), 27 | dict(type='RandomFlip', flip_ratio=0.5), 28 | dict(type='Normalize', **img_norm_cfg), 29 | dict(type='Pad', size_divisor=32), 30 | dict(type='SegRescale', scale_factor=1 / 8), 31 | dict(type='DefaultFormatBundle'), 32 | dict( 33 | type='Collect', 34 | keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks', 'gt_semantic_seg']), 35 | ] 36 | test_pipeline = [ 37 | dict(type='LoadImageFromFile'), 38 | dict( 39 | type='MultiScaleFlipAug', 40 | img_scale=(1333, 800), 41 | flip=False, 42 | transforms=[ 43 | dict(type='Resize', keep_ratio=True), 44 | dict(type='RandomFlip', flip_ratio=0.5), 45 | dict(type='Normalize', **img_norm_cfg), 46 | dict(type='Pad', size_divisor=32), 47 | dict(type='ImageToTensor', keys=['img']), 48 | dict(type='Collect', keys=['img']), 49 | ]) 50 | ] 51 | data = dict( 52 | train=dict( 53 | seg_prefix=data_root + 'stuffthingmaps/train2017/', 54 | pipeline=train_pipeline), 55 | val=dict(pipeline=test_pipeline), 56 | test=dict(pipeline=test_pipeline)) 57 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/htc/zzd_htc_without_semantic_r50_fpn_1x_crosswalk.py: -------------------------------------------------------------------------------- 1 | _base_ = [ 2 | '../_base_/datasets/zzd_crosswalk_instance.py', 3 | '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py' 4 | ] 5 | # model settings 6 | model = dict( 7 | type='HybridTaskCascade', 8 | backbone=dict( 9 | type='ResNet', 10 | depth=50, 11 | num_stages=4, 12 | out_indices=(0, 1, 2, 3), 13 | frozen_stages=1, 14 | norm_cfg=dict(type='BN', requires_grad=True), 15 | norm_eval=True, 16 | style='pytorch', 17 | init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50')), 18 | neck=dict( 19 | type='FPN', 20 | in_channels=[256, 512, 1024, 2048], 21 | out_channels=256, 22 | num_outs=5), 23 | rpn_head=dict( 24 | type='RPNHead', 25 | in_channels=256, 26 | feat_channels=256, 27 | anchor_generator=dict( 28 | type='AnchorGenerator', 29 | scales=[8], 30 | ratios=[0.5, 1.0, 2.0], 31 | strides=[4, 8, 16, 32, 64]), 32 | bbox_coder=dict( 33 | type='DeltaXYWHBBoxCoder', 34 | target_means=[.0, .0, .0, .0], 35 | target_stds=[1.0, 1.0, 1.0, 1.0]), 36 | loss_cls=dict( 37 | type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0), 38 | loss_bbox=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0)), 39 | roi_head=dict( 40 | type='HybridTaskCascadeRoIHead', 41 | interleaved=True, 42 | mask_info_flow=True, 43 | num_stages=3, 44 | stage_loss_weights=[1, 0.5, 0.25], 45 | bbox_roi_extractor=dict( 46 | type='SingleRoIExtractor', 47 | roi_layer=dict(type='RoIAlign', output_size=7, sampling_ratio=0), 48 | out_channels=256, 49 | featmap_strides=[4, 8, 16, 32]), 50 | bbox_head=[ 51 | dict( 52 | type='Shared2FCBBoxHead', 53 | in_channels=256, 54 | fc_out_channels=1024, 55 | roi_feat_size=7, 56 | num_classes=2, 57 | bbox_coder=dict( 58 | type='DeltaXYWHBBoxCoder', 59 | target_means=[0., 0., 0., 0.], 60 | target_stds=[0.1, 0.1, 0.2, 0.2]), 61 | reg_class_agnostic=True, 62 | loss_cls=dict( 63 | type='CrossEntropyLoss', 64 | use_sigmoid=False, 65 | loss_weight=1.0), 66 | loss_bbox=dict(type='SmoothL1Loss', beta=1.0, 67 | loss_weight=1.0)), 68 | dict( 69 | type='Shared2FCBBoxHead', 70 | in_channels=256, 71 | fc_out_channels=1024, 72 | roi_feat_size=7, 73 | num_classes=2, 74 | bbox_coder=dict( 75 | type='DeltaXYWHBBoxCoder', 76 | target_means=[0., 0., 0., 0.], 77 | target_stds=[0.05, 0.05, 0.1, 0.1]), 78 | reg_class_agnostic=True, 79 | loss_cls=dict( 80 | type='CrossEntropyLoss', 81 | use_sigmoid=False, 82 | loss_weight=1.0), 83 | loss_bbox=dict(type='SmoothL1Loss', beta=1.0, 84 | loss_weight=1.0)), 85 | dict( 86 | type='Shared2FCBBoxHead', 87 | in_channels=256, 88 | fc_out_channels=1024, 89 | roi_feat_size=7, 90 | num_classes=2, 91 | bbox_coder=dict( 92 | type='DeltaXYWHBBoxCoder', 93 | target_means=[0., 0., 0., 0.], 94 | target_stds=[0.033, 0.033, 0.067, 0.067]), 95 | reg_class_agnostic=True, 96 | loss_cls=dict( 97 | type='CrossEntropyLoss', 98 | use_sigmoid=False, 99 | loss_weight=1.0), 100 | loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)) 101 | ], 102 | mask_roi_extractor=dict( 103 | type='SingleRoIExtractor', 104 | roi_layer=dict(type='RoIAlign', output_size=14, sampling_ratio=0), 105 | out_channels=256, 106 | featmap_strides=[4, 8, 16, 32]), 107 | mask_head=[ 108 | dict( 109 | type='HTCMaskHead', 110 | with_conv_res=False, 111 | num_convs=4, 112 | in_channels=256, 113 | conv_out_channels=256, 114 | num_classes=2, 115 | loss_mask=dict( 116 | type='CrossEntropyLoss', use_mask=True, loss_weight=1.0)), 117 | dict( 118 | type='HTCMaskHead', 119 | num_convs=4, 120 | in_channels=256, 121 | conv_out_channels=256, 122 | num_classes=2, 123 | loss_mask=dict( 124 | type='CrossEntropyLoss', use_mask=True, loss_weight=1.0)), 125 | dict( 126 | type='HTCMaskHead', 127 | num_convs=4, 128 | in_channels=256, 129 | conv_out_channels=256, 130 | num_classes=2, 131 | loss_mask=dict( 132 | type='CrossEntropyLoss', use_mask=True, loss_weight=1.0)) 133 | ]), 134 | # model training and testing settings 135 | train_cfg=dict( 136 | rpn=dict( 137 | assigner=dict( 138 | type='MaxIoUAssigner', 139 | pos_iou_thr=0.7, 140 | neg_iou_thr=0.3, 141 | min_pos_iou=0.3, 142 | ignore_iof_thr=-1), 143 | sampler=dict( 144 | type='RandomSampler', 145 | num=256, 146 | pos_fraction=0.5, 147 | neg_pos_ub=-1, 148 | add_gt_as_proposals=False), 149 | allowed_border=0, 150 | pos_weight=-1, 151 | debug=False), 152 | rpn_proposal=dict( 153 | nms_pre=2000, 154 | max_per_img=2000, 155 | nms=dict(type='nms', iou_threshold=0.7), 156 | min_bbox_size=0), 157 | rcnn=[ 158 | dict( 159 | assigner=dict( 160 | type='MaxIoUAssigner', 161 | pos_iou_thr=0.5, 162 | neg_iou_thr=0.5, 163 | min_pos_iou=0.5, 164 | ignore_iof_thr=-1), 165 | sampler=dict( 166 | type='RandomSampler', 167 | num=512, 168 | pos_fraction=0.25, 169 | neg_pos_ub=-1, 170 | add_gt_as_proposals=True), 171 | mask_size=28, 172 | pos_weight=-1, 173 | debug=False), 174 | dict( 175 | assigner=dict( 176 | type='MaxIoUAssigner', 177 | pos_iou_thr=0.6, 178 | neg_iou_thr=0.6, 179 | min_pos_iou=0.6, 180 | ignore_iof_thr=-1), 181 | sampler=dict( 182 | type='RandomSampler', 183 | num=512, 184 | pos_fraction=0.25, 185 | neg_pos_ub=-1, 186 | add_gt_as_proposals=True), 187 | mask_size=28, 188 | pos_weight=-1, 189 | debug=False), 190 | dict( 191 | assigner=dict( 192 | type='MaxIoUAssigner', 193 | pos_iou_thr=0.7, 194 | neg_iou_thr=0.7, 195 | min_pos_iou=0.7, 196 | ignore_iof_thr=-1), 197 | sampler=dict( 198 | type='RandomSampler', 199 | num=512, 200 | pos_fraction=0.25, 201 | neg_pos_ub=-1, 202 | add_gt_as_proposals=True), 203 | mask_size=28, 204 | pos_weight=-1, 205 | debug=False) 206 | ]), 207 | test_cfg=dict( 208 | rpn=dict( 209 | nms_pre=1000, 210 | max_per_img=1000, 211 | nms=dict(type='nms', iou_threshold=0.7), 212 | min_bbox_size=0), 213 | rcnn=dict( 214 | score_thr=0.001, 215 | nms=dict(type='nms', iou_threshold=0.5), 216 | max_per_img=100, 217 | mask_thr_binary=0.5))) 218 | img_norm_cfg = dict( 219 | mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) 220 | test_pipeline = [ 221 | dict(type='LoadImageFromFile'), 222 | dict( 223 | type='MultiScaleFlipAug', 224 | img_scale=(1333, 800), 225 | flip=False, 226 | transforms=[ 227 | dict(type='Resize', keep_ratio=True), 228 | dict(type='RandomFlip', flip_ratio=0.5), 229 | dict(type='Normalize', **img_norm_cfg), 230 | dict(type='Pad', size_divisor=32), 231 | dict(type='ImageToTensor', keys=['img']), 232 | dict(type='Collect', keys=['img']), 233 | ]) 234 | ] 235 | data = dict( 236 | val=dict(pipeline=test_pipeline), test=dict(pipeline=test_pipeline)) 237 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/mask_rcnn/zzd_mask_rcnn_r50_fpn_1x_coco.py: -------------------------------------------------------------------------------- 1 | _base_ = [ 2 | '../_base_/models/mask_rcnn_r50_fpn.py', 3 | '../_base_/datasets/zzd_coco_instance.py', 4 | '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py' 5 | ] 6 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/mask_rcnn/zzd_mask_rcnn_r50_fpn_1x_crosswalk.py: -------------------------------------------------------------------------------- 1 | _base_ = [ 2 | '../_base_/models/mask_rcnn_r50_fpn.py', 3 | '../_base_/datasets/zzd_crosswalk_instance.py', 4 | '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py' 5 | ] 6 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/scnet/zzd_scnet_r50_fpn_1x_crosswalk.py: -------------------------------------------------------------------------------- 1 | _base_ = '../htc/zzd_htc_r50_fpn_1x_crosswalk.py' 2 | # model settings 3 | model = dict( 4 | type='SCNet', 5 | roi_head=dict( 6 | _delete_=True, 7 | type='SCNetRoIHead', 8 | num_stages=3, 9 | stage_loss_weights=[1, 0.5, 0.25], 10 | bbox_roi_extractor=dict( 11 | type='SingleRoIExtractor', 12 | roi_layer=dict(type='RoIAlign', output_size=7, sampling_ratio=0), 13 | out_channels=256, 14 | featmap_strides=[4, 8, 16, 32]), 15 | bbox_head=[ 16 | dict( 17 | type='SCNetBBoxHead', 18 | num_shared_fcs=2, 19 | in_channels=256, 20 | fc_out_channels=1024, 21 | roi_feat_size=7, 22 | num_classes=2, 23 | bbox_coder=dict( 24 | type='DeltaXYWHBBoxCoder', 25 | target_means=[0., 0., 0., 0.], 26 | target_stds=[0.1, 0.1, 0.2, 0.2]), 27 | reg_class_agnostic=True, 28 | loss_cls=dict( 29 | type='CrossEntropyLoss', 30 | use_sigmoid=False, 31 | loss_weight=1.0), 32 | loss_bbox=dict(type='SmoothL1Loss', beta=1.0, 33 | loss_weight=1.0)), 34 | dict( 35 | type='SCNetBBoxHead', 36 | num_shared_fcs=2, 37 | in_channels=256, 38 | fc_out_channels=1024, 39 | roi_feat_size=7, 40 | num_classes=2, 41 | bbox_coder=dict( 42 | type='DeltaXYWHBBoxCoder', 43 | target_means=[0., 0., 0., 0.], 44 | target_stds=[0.05, 0.05, 0.1, 0.1]), 45 | reg_class_agnostic=True, 46 | loss_cls=dict( 47 | type='CrossEntropyLoss', 48 | use_sigmoid=False, 49 | loss_weight=1.0), 50 | loss_bbox=dict(type='SmoothL1Loss', beta=1.0, 51 | loss_weight=1.0)), 52 | dict( 53 | type='SCNetBBoxHead', 54 | num_shared_fcs=2, 55 | in_channels=256, 56 | fc_out_channels=1024, 57 | roi_feat_size=7, 58 | num_classes=2, 59 | bbox_coder=dict( 60 | type='DeltaXYWHBBoxCoder', 61 | target_means=[0., 0., 0., 0.], 62 | target_stds=[0.033, 0.033, 0.067, 0.067]), 63 | reg_class_agnostic=True, 64 | loss_cls=dict( 65 | type='CrossEntropyLoss', 66 | use_sigmoid=False, 67 | loss_weight=1.0), 68 | loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)) 69 | ], 70 | mask_roi_extractor=dict( 71 | type='SingleRoIExtractor', 72 | roi_layer=dict(type='RoIAlign', output_size=14, sampling_ratio=0), 73 | out_channels=256, 74 | featmap_strides=[4, 8, 16, 32]), 75 | mask_head=dict( 76 | type='SCNetMaskHead', 77 | num_convs=12, 78 | in_channels=256, 79 | conv_out_channels=256, 80 | num_classes=2, 81 | conv_to_res=True, 82 | loss_mask=dict( 83 | type='CrossEntropyLoss', use_mask=True, loss_weight=1.0)), 84 | semantic_roi_extractor=dict( 85 | type='SingleRoIExtractor', 86 | roi_layer=dict(type='RoIAlign', output_size=14, sampling_ratio=0), 87 | out_channels=256, 88 | featmap_strides=[8]), 89 | semantic_head=dict( 90 | type='SCNetSemanticHead', 91 | num_ins=5, 92 | fusion_level=1, 93 | num_convs=4, 94 | in_channels=256, 95 | conv_out_channels=256, 96 | num_classes=183, 97 | loss_seg=dict( 98 | type='CrossEntropyLoss', ignore_index=255, loss_weight=0.2), 99 | conv_to_res=True), 100 | glbctx_head=dict( 101 | type='GlobalContextHead', 102 | num_convs=4, 103 | in_channels=256, 104 | conv_out_channels=256, 105 | num_classes=2, 106 | loss_weight=3.0, 107 | conv_to_res=True), 108 | feat_relay_head=dict( 109 | type='FeatureRelayHead', 110 | in_channels=1024, 111 | out_conv_channels=256, 112 | roi_feat_size=7, 113 | scale_factor=2))) 114 | 115 | # uncomment below code to enable test time augmentations 116 | # img_norm_cfg = dict( 117 | # mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) 118 | # test_pipeline = [ 119 | # dict(type='LoadImageFromFile'), 120 | # dict( 121 | # type='MultiScaleFlipAug', 122 | # img_scale=[(600, 900), (800, 1200), (1000, 1500), (1200, 1800), 123 | # (1400, 2100)], 124 | # flip=True, 125 | # transforms=[ 126 | # dict(type='Resize', keep_ratio=True), 127 | # dict(type='RandomFlip', flip_ratio=0.5), 128 | # dict(type='Normalize', **img_norm_cfg), 129 | # dict(type='Pad', size_divisor=32), 130 | # dict(type='ImageToTensor', keys=['img']), 131 | # dict(type='Collect', keys=['img']), 132 | # ]) 133 | # ] 134 | # data = dict( 135 | # val=dict(pipeline=test_pipeline), 136 | # test=dict(pipeline=test_pipeline)) 137 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/swin/zzd_mask_rcnn_swin-t-p4-w7_fpn_1x_crosswalk.py: -------------------------------------------------------------------------------- 1 | _base_ = [ 2 | '../_base_/models/mask_rcnn_r50_fpn.py', 3 | '../_base_/datasets/zzd_crosswalk_instance.py', 4 | '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py' 5 | ] 6 | pretrained = 'https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_tiny_patch4_window7_224.pth' # noqa 7 | # pretrained = None 8 | model = dict( 9 | type='MaskRCNN', 10 | backbone=dict( 11 | _delete_=True, 12 | type='SwinTransformer', 13 | embed_dims=96, 14 | depths=[2, 2, 6, 2], 15 | num_heads=[3, 6, 12, 24], 16 | window_size=7, 17 | mlp_ratio=4, 18 | qkv_bias=True, 19 | qk_scale=None, 20 | drop_rate=0., 21 | attn_drop_rate=0., 22 | drop_path_rate=0.2, 23 | patch_norm=True, 24 | out_indices=(0, 1, 2, 3), 25 | with_cp=False, 26 | convert_weights=True, 27 | init_cfg=dict(type='Pretrained', checkpoint=pretrained)), 28 | neck=dict(in_channels=[96, 192, 384, 768])) 29 | 30 | optimizer = dict( 31 | _delete_=True, 32 | type='AdamW', 33 | lr=0.0001, 34 | betas=(0.9, 0.999), 35 | weight_decay=0.05, 36 | paramwise_cfg=dict( 37 | custom_keys={ 38 | 'absolute_pos_embed': dict(decay_mult=0.), 39 | 'relative_position_bias_table': dict(decay_mult=0.), 40 | 'norm': dict(decay_mult=0.) 41 | })) 42 | lr_config = dict(warmup_iters=1000, step=[8, 11]) 43 | runner = dict(max_epochs=12) 44 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/yolox/zzd_yolox_m_8x8_300e_crosswalk.py: -------------------------------------------------------------------------------- 1 | _base_ = './zzd_yolox_s_8x8_300e_crosswalk.py' 2 | 3 | # model settings 4 | model = dict( 5 | backbone=dict(deepen_factor=0.67, widen_factor=0.75), 6 | neck=dict(in_channels=[192, 384, 768], out_channels=192, num_csp_blocks=2), 7 | bbox_head=dict(in_channels=192, feat_channels=192), 8 | ) 9 | -------------------------------------------------------------------------------- /repos/mmdetection/configs/yolox/zzd_yolox_s_8x8_300e_crosswalk.py: -------------------------------------------------------------------------------- 1 | _base_ = ['../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'] 2 | 3 | # model settings 4 | model = dict( 5 | type='YOLOX', 6 | backbone=dict(type='CSPDarknet', deepen_factor=0.33, widen_factor=0.5), 7 | neck=dict( 8 | type='YOLOXPAFPN', 9 | in_channels=[128, 256, 512], 10 | out_channels=128, 11 | num_csp_blocks=1), 12 | bbox_head=dict( 13 | type='YOLOXHead', num_classes=2, in_channels=128, feat_channels=128), 14 | train_cfg=dict(assigner=dict(type='SimOTAAssigner', center_radius=2.5)), 15 | # In order to align the source code, the threshold of the val phase is 16 | # 0.01, and the threshold of the test phase is 0.001. 17 | test_cfg=dict(score_thr=0.01, nms=dict(type='nms', iou_threshold=0.65))) 18 | 19 | # dataset settings 20 | # data_root = 'data/coco/' 21 | dataset_type = 'CocoDataset' 22 | classes = ('crosswalk', 'guide_arrows') 23 | data_root = '/home/zzd/datasets/crosswalk/fogged_train_data_coco_format/' 24 | 25 | img_norm_cfg = dict( 26 | mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) 27 | 28 | img_scale = (640, 640) 29 | 30 | train_pipeline = [ 31 | dict(type='Mosaic', img_scale=img_scale, pad_val=114.0), 32 | dict( 33 | type='RandomAffine', 34 | scaling_ratio_range=(0.1, 2), 35 | border=(-img_scale[0] // 2, -img_scale[1] // 2)), 36 | dict( 37 | type='MixUp', 38 | img_scale=img_scale, 39 | ratio_range=(0.8, 1.6), 40 | pad_val=114.0), 41 | dict( 42 | type='PhotoMetricDistortion', 43 | brightness_delta=32, 44 | contrast_range=(0.5, 1.5), 45 | saturation_range=(0.5, 1.5), 46 | hue_delta=18), 47 | dict(type='RandomFlip', flip_ratio=0.5), 48 | dict(type='Resize', keep_ratio=True), 49 | dict(type='Pad', pad_to_square=True, pad_val=114.0), 50 | dict(type='Normalize', **img_norm_cfg), 51 | dict(type='DefaultFormatBundle'), 52 | dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']) 53 | ] 54 | 55 | train_dataset = dict( 56 | type='MultiImageMixDataset', 57 | dataset=dict( 58 | type=dataset_type, 59 | ann_file=data_root + 'annotations/instances_train2017.json', 60 | img_prefix=data_root + 'train2017/', 61 | classes=classes, 62 | pipeline=[ 63 | dict(type='LoadImageFromFile', to_float32=True), 64 | dict(type='LoadAnnotations', with_bbox=True) 65 | ], 66 | filter_empty_gt=False, 67 | ), 68 | pipeline=train_pipeline, 69 | dynamic_scale=img_scale) 70 | 71 | test_pipeline = [ 72 | dict(type='LoadImageFromFile'), 73 | dict( 74 | type='MultiScaleFlipAug', 75 | img_scale=img_scale, 76 | flip=False, 77 | transforms=[ 78 | dict(type='Resize', keep_ratio=True), 79 | dict(type='RandomFlip'), 80 | dict(type='Pad', size=img_scale, pad_val=114.0), 81 | dict(type='Normalize', **img_norm_cfg), 82 | dict(type='DefaultFormatBundle'), 83 | dict(type='Collect', keys=['img']) 84 | ]) 85 | ] 86 | 87 | data = dict( 88 | samples_per_gpu=4, 89 | workers_per_gpu=2, 90 | train=train_dataset, 91 | val=dict( 92 | type=dataset_type, 93 | ann_file=data_root + 'annotations/instances_val2017.json', 94 | img_prefix=data_root + 'val2017/', 95 | classes=classes, 96 | pipeline=test_pipeline), 97 | test=dict( 98 | type=dataset_type, 99 | ann_file=data_root + 'annotations/instances_val2017.json', 100 | img_prefix=data_root + 'val2017/', 101 | classes=classes, 102 | pipeline=test_pipeline)) 103 | 104 | # optimizer 105 | # default 8 gpu 106 | optimizer = dict( 107 | type='SGD', 108 | lr=0.01, 109 | momentum=0.9, 110 | weight_decay=5e-4, 111 | nesterov=True, 112 | paramwise_cfg=dict(norm_decay_mult=0., bias_decay_mult=0.)) 113 | optimizer_config = dict(grad_clip=None) 114 | 115 | # learning policy 116 | lr_config = dict( 117 | _delete_=True, 118 | policy='YOLOX', 119 | warmup='exp', 120 | by_epoch=False, 121 | warmup_by_epoch=True, 122 | warmup_ratio=1, 123 | warmup_iters=5, # 5 epoch 124 | num_last_epochs=15, 125 | min_lr_ratio=0.05) 126 | runner = dict(type='EpochBasedRunner', max_epochs=100) 127 | 128 | resume_from = None 129 | interval = 10 130 | 131 | custom_hooks = [ 132 | dict(type='YOLOXModeSwitchHook', num_last_epochs=15, priority=48), 133 | dict( 134 | type='SyncRandomSizeHook', 135 | ratio_range=(14, 26), 136 | img_scale=img_scale, 137 | priority=48), 138 | dict( 139 | type='SyncNormHook', 140 | num_last_epochs=15, 141 | interval=interval, 142 | priority=48), 143 | dict(type='ExpMomentumEMAHook', resume_from=resume_from, priority=49) 144 | ] 145 | checkpoint_config = dict(interval=interval) 146 | evaluation = dict(interval=interval, metric='bbox') 147 | log_config = dict(interval=50) 148 | -------------------------------------------------------------------------------- /repos/mmdetection/demo/zzd_inference_speed_test.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) OpenMMLab. All rights reserved. 2 | import os 3 | import asyncio 4 | import numpy as np 5 | from pathlib import Path 6 | from argparse import ArgumentParser 7 | 8 | from mmdet.apis import (async_inference_detector, inference_detector, 9 | init_detector, show_result_pyplot) 10 | 11 | 12 | def parse_args(): 13 | parser = ArgumentParser() 14 | parser.add_argument('--img', help='Image file') 15 | parser.add_argument('--config', help='Config file') 16 | parser.add_argument('--checkpoint', help='Checkpoint file') 17 | parser.add_argument( 18 | '--device', default='cuda:0', help='Device used for inference') 19 | parser.add_argument( 20 | '--score-thr', type=float, default=0.3, help='bbox score threshold') 21 | parser.add_argument( 22 | '--async-test', 23 | action='store_true', 24 | help='whether to set async options for async inference.') 25 | args = parser.parse_args() 26 | 27 | path = '../../example/images' 28 | imgs = sorted([f'{path}/{x}' for x in os.listdir(path) if x.endswith('jpg')]) 29 | args.img = imgs 30 | 31 | ck_dict = dict(mask_rcnn='zzd_mask_rcnn_r50_fpn_1x_crosswalk/epoch_12.pth', 32 | faster_rcnn='zzd_faster_rcnn_r50_fpn_1x_crosswalk/epoch_12.pth', 33 | yolox='zzd_yolox_m_8x8_300e_crosswalk/epoch_100.pth', 34 | swin='zzd_mask_rcnn_swin-t-p4-w7_fpn_1x_crosswalk/epoch_12.pth') 35 | 36 | model = 'mask_rcnn' 37 | model = 'faster_rcnn' 38 | model = 'yolox' 39 | model = 'swin' 40 | args.config = f'./configs/{model}/{ck_dict[model].split("/")[0]}.py' 41 | args.checkpoint = f'work_dirs/{ck_dict[model]}' 42 | return args 43 | 44 | 45 | def main(args): 46 | # build the model from a config file and a checkpoint file 47 | model = init_detector(args.config, args.checkpoint, device=args.device) 48 | # test a single image 49 | t = [] 50 | for i, img in enumerate(args.img): 51 | result, st = inference_detector(model, img) # spent time 52 | print(f'{i:>2} spent time: {st}') 53 | if i != 0: 54 | t.append(st) 55 | t = np.array(t) 56 | print(f'total time : {np.sum(t):.4f}s mean: {np.mean(t)*1000:.2f}ms FPS: {1/np.mean(t):.2f}') 57 | # show the results 58 | # show_result_pyplot(model, args.img, result, score_thr=args.score_thr) 59 | 60 | 61 | async def async_main(args): 62 | # build the model from a config file and a checkpoint file 63 | model = init_detector(args.config, args.checkpoint, device=args.device) 64 | # test a single image 65 | tasks = asyncio.create_task(async_inference_detector(model, args.img)) 66 | result = await asyncio.gather(tasks) 67 | # show the results 68 | show_result_pyplot(model, args.img, result[0], score_thr=args.score_thr) 69 | 70 | 71 | if __name__ == '__main__': 72 | args = parse_args() 73 | if args.async_test: 74 | asyncio.run(async_main(args)) 75 | else: 76 | main(args) 77 | -------------------------------------------------------------------------------- /repos/mmdetection/zzd_inference.py: -------------------------------------------------------------------------------- 1 | import os 2 | import subprocess 3 | 4 | img = 'demo/demo.jpg' 5 | # cfg = './configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py' 6 | # checkpoint = './checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth' 7 | 8 | img = '../../example/images/filename_00038.jpg' 9 | cfg = './configs/mask_rcnn/zzd_mask_rcnn_r50_fpn_1x_crosswalk.py' 10 | checkpoint = 'work_dirs/zzd_mask_rcnn_r50_fpn_1x_crosswalk/epoch_12.pth' 11 | 12 | code = f'python demo/image_demo.py {img}' \ 13 | f' {cfg} {checkpoint} --device cpu' 14 | print(code) 15 | # subprocess.call(code) 16 | os.system(code) 17 | -------------------------------------------------------------------------------- /repos/mmdetection/zzd_train.py: -------------------------------------------------------------------------------- 1 | import os 2 | import subprocess 3 | 4 | cfg = "./configs/mask_rcnn/zzd_mask_rcnn_r50_fpn_1x_crosswalk.py" 5 | cfg = "./configs/faster_rcnn/zzd_faster_rcnn_r50_fpn_1x_crosswalk.py" 6 | cfg = "./configs/centernet/zzd_centernet_resnet18_dcnv2_140e_crosswalk.py" 7 | cfg = './configs/swin/zzd_mask_rcnn_swin-t-p4-w7_fpn_1x_crosswalk.py' 8 | cfg = './configs/yolox/zzd_yolox_m_8x8_300e_crosswalk.py' 9 | 10 | code = f'python tools/train.py {cfg}' 11 | # subprocess.call(code) 12 | os.system(code) 13 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | # pip install -r requirements.txt 2 | 3 | Cython 4 | matplotlib>=3.2.2 5 | numpy>=1.18.5 6 | opencv-python>=3.4.2 7 | Pillow 8 | pytorch>=1.6.0 9 | torchvision>=0.7.0 10 | PyYAML>=5.3 11 | scipy>=1.4.1 12 | tensorboard>=2.2 13 | torch>=1.7.0 14 | torchvision>=0.8.1 15 | tqdm>=4.41.0 16 | opencv_python 17 | 18 | -------------------------------------------------------------------------------- /scripts/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/scripts/__init__.py -------------------------------------------------------------------------------- /scripts/check_coco.py: -------------------------------------------------------------------------------- 1 | """ 2 | 检查coco数据集 3 | """ 4 | import os, sys 5 | import argparse 6 | 7 | sys.path.append(f'/home/zzd/PycharmProject/damei') 8 | import damei as dm 9 | 10 | # parser = argparse.ArgumentParser() 11 | # parser.add_argument('--json_path', ) 12 | 13 | 14 | dp = "/home/zzd/datasets/crosswalk/fogged_train_data_coco_format" 15 | jp = f'{dp}/annotations/instances_train2017.json' 16 | # jp = "/home/zzd/datasets/mscoco/annotations/instances_train2017.json" 17 | dm.tools.check_coco(json_path=jp) 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | -------------------------------------------------------------------------------- /scripts/check_dataset.py: -------------------------------------------------------------------------------- 1 | """ 2 | Check datasets 3 | """ 4 | import os, sys 5 | import numpy as np 6 | import cv2 7 | from pathlib import Path 8 | import random 9 | import argparse 10 | 11 | try: 12 | from utils.datasets import LoadImages, LoadImagesAndLabels 13 | from utils.general import plot_one_box, xywh2xyxy 14 | except: 15 | sys.path.append('..') 16 | from utils.datasets import LoadImages, LoadImagesAndLabels 17 | from utils.general import plot_one_box, xywh2xyxy 18 | 19 | 20 | class CheckDataset(object): 21 | def __init__(self, dp): 22 | self.dp = dp 23 | # self.colors = [[random.randint(0, 255) for _ in range(3)] for __ in range(100)] 24 | # self.colors = ['#009966', '#EE3B3B'] 25 | self.colors = [[102, 153, 0], [59, 59, 238]] 26 | self.names = ['crosswalk', 'guide_arrows'] 27 | 28 | def __call__(self, trte='test', cf=False): 29 | p = f'{self.dp}/images/{trte}' 30 | 31 | imgs = [f'{self.dp}/images/{trte}/{x}' for x in os.listdir(p) if x.endswith('.jpg')] 32 | imgs = sorted(imgs) 33 | for i, imgp in enumerate(imgs): 34 | stem = Path(imgp).stem 35 | labelp = f'{self.dp}/labels/{trte}/{Path(imgp).stem}.txt' 36 | img = cv2.imread(imgp) 37 | h, w, c = img.shape 38 | 39 | with open(labelp, 'r') as f: 40 | label = f.readlines() 41 | label = np.array([x.split() for x in label], dtype=np.float32) 42 | 43 | classes = label[:, 0] 44 | bboxes = label[:, 1::] 45 | bboxes = xywh2xyxy(bboxes) 46 | for j in range(len(label)): 47 | cls = classes[j] 48 | bbox = bboxes[j] 49 | bbox[0] *= w 50 | bbox[1] *= h 51 | bbox[2] *= w 52 | bbox[3] *= h 53 | plot_one_box(bbox, img, label=f'{self.names[int(cls)]}', color=self.colors[int(cls)]) 54 | 55 | print(f'imgs: {len(imgs)} stem: {stem} img_shape: {img.shape} lb: {label}') 56 | # cr = np.any(label[:, 0] == 1) 57 | crit = 'fogged' in stem if cf else True 58 | if crit: 59 | cv2.imshow('xx', img) 60 | if cv2.waitKey(0) == ord('q'): 61 | exit() 62 | 63 | 64 | if __name__ == '__main__': 65 | parser = argparse.ArgumentParser() 66 | parser.add_argument('-dp', '--dataset-path', type=str, 67 | help='path to dataset need to be checked. such as: xxx/train_data_v5_format', 68 | default='/home/zzd/datasets/crosswalk/train_data_v5_format') 69 | parser.add_argument('-trte', '--train_or_test', type=str, 70 | help='specify check train or test of YOLOv5 format', 71 | default='train') 72 | parser.add_argument('-cf', '--only-check-fog', action='store_true', help='only check fogged dataset') 73 | 74 | opt = parser.parse_args() 75 | cd = CheckDataset(dp=opt.dataset_path) 76 | cd(trte=opt.train_or_test, cf=opt.only_check_fog) 77 | -------------------------------------------------------------------------------- /scripts/get_confusing_matrix.py: -------------------------------------------------------------------------------- 1 | """ 2 | 针对crosswalk 1770测试集，计算混淆矩阵 3 | """ 4 | 5 | import os 6 | import numpy as np 7 | 8 | import time 9 | from PIL import Image 10 | from pathlib import Path 11 | import re 12 | 13 | 14 | class GetConfusingMatrix(object): 15 | """ 16 | for crosswalk detection. 17 | """ 18 | def __init__(self, thresh=0.8, img_suffix='jpg'): 19 | self.thresh = thresh 20 | self.img_suffix = img_suffix 21 | self.classes = ['crosswalk'] 22 | 23 | def analyse(self, thresh=None, v5_out=None, pn_dir=None): 24 | print(f"{Path(v5_out).stem:-^80}") 25 | 26 | self.thresh = thresh if thresh is not None else self.thresh 27 | # 获取detect data 28 | dt_data = self.get_detect_data(v5_out=v5_out) 29 | gt_data = self.get_ground_truth(pn_dir, num_dt=dt_data.shape[0]) # array [n, 2] 2:filestem, cls, cls0: 无斑马线，cls1: 有斑马线 30 | assert dt_data.shape[0] == gt_data.shape[0], \ 31 | f'dt shape: {dt_data.shape[0]} is not compatible with gt shape: {gt_data.shape[0]}' 32 | # array [n_files, 100, 3] 100: 同一张图最大检测到的目标数目，3: filename, cls, conf 33 | assert len(self.classes) == 1 34 | for i, cls in enumerate(self.classes): 35 | # gt = gt_data[gt_data[:, 1] == str(i)] 36 | dt = np.array(dt_data[np.where(dt_data[:, :, 1] == str(i))], dtype=np.str_) # (n, 7) # 筛选出该类 37 | dt = dt[np.array(dt[:, 2], dtype=np.float32) > self.thresh] 38 | # new_dt = np.zeros((dt_data.shape[0], dt_data.shape[2])).astype(np.str) # (1770, 7) 39 | ret = self.cal_TPFPTNFN(cal_cls=self.classes[i], dt_data=dt, gt_data=gt_data) 40 | return ret 41 | 42 | def cal_TPFPTNFN(self, cal_cls, dt_data, gt_data): 43 | """ 44 | 计算混淆矩阵。 45 | :param cal_cls: 需要计算的类别名称 46 | :param dt_data: 检测结果：[n, 7] n: 所有检测的图，7: img, cls, conf, xc, yc, w, h 47 | :param gt_data: 真值 48 | :return: 49 | """ 50 | TP_count, FP_count, TN_count, FN_count = 0, 0, 0, 0 51 | 52 | dt_filenames = np.unique(dt_data[:, 0]).tolist() 53 | for i, (filename, gtcls) in enumerate(gt_data): # 54 | gtcls = int(gtcls) 55 | if gtcls == 1: # 1是有斑马线 56 | gt = True 57 | else: 58 | gt = False 59 | 60 | # filename, cls, conf = dt_data[i, :3] 61 | if filename in dt_filenames: # 检测到了，有斑马线 62 | dt = True 63 | else: 64 | dt = False 65 | 66 | if gt and dt: 67 | TP_count += 1 68 | elif gt and not dt: 69 | FN_count += 1 70 | elif not gt and dt: 71 | FP_count += 1 72 | elif not gt and not dt: 73 | TN_count += 1 74 | else: 75 | raise Exception(f'ERROR, gt:{gt} dt:{dt}') 76 | # print(filename, gt, dt) 77 | # if i > 100: 78 | # break 79 | confusing_matrix = np.array([TP_count, FP_count, FN_count, TN_count]).astype(np.int32) 80 | # print(confusing_matrix) 81 | # exit() 82 | 83 | # confusing_matrix = confusing_matrix/len(data) 84 | # print(confusing_matrix) 85 | alpha = 1e-10 86 | 87 | ACC = (TP_count + TN_count) / (TP_count + TN_count + FP_count + FN_count+alpha) # 准确率 88 | PPV = (TP_count) / (TP_count + FP_count+alpha) # 精确率 89 | TPR = (TP_count) / (TP_count + FN_count+alpha) # 召回率 90 | TNR = (TN_count) / (TN_count + FP_count+alpha) # 特异度 91 | F1_score = 2 * PPV * TPR / (PPV + TPR+alpha) 92 | 93 | cm = confusing_matrix 94 | string1 = f"{'cls':^15}{'ALL':<5}{'TP':<5}{'FP':<5}{'FN':<5}{'TN':<5}\n" \ 95 | f"{cal_cls:^15}{len(dt_data):<5}{cm[0]:<5}{cm[1]:<5}{cm[2]:<5}{cm[3]:<5}" 96 | # print(f"ALL TP FP FN TN") 97 | # print(f"{len(dt_data)} {confusing_matrix}") 98 | print(string1) 99 | string2 = f'Thresh: {self.thresh} Accuracy: {ACC*100:.2f}% Precision: {PPV*100:.2f}% Recall: {TPR*100:.2f}% ' \ 100 | f'Specificity: {TNR*100:.2f}% F1_score: {F1_score*100:.2f}%' 101 | print(string2) 102 | pnts3 = f"{len(dt_data):<5}{cm[0]:<5}{cm[1]:<5}{cm[2]:<5}{cm[3]:<5} {self.thresh} {ACC*100:.2f} " \ 103 | f"{PPV*100:.2f} {TPR*100:.2f} {TNR*100:.2f} {F1_score*100:.2f}" 104 | print(pnts3) 105 | return pnts3 106 | 107 | def get_ground_truth(self, pn_dir, num_dt): 108 | with open(f'{pn_dir}/positive.txt', 'r') as f: 109 | positive = f.readlines() 110 | with open(f'{pn_dir}/negative.txt', 'r') as f: 111 | negative = f.readlines() 112 | positive = [x.strip('\n') for x in positive] 113 | negative = [x.strip('\n') for x in negative] 114 | gt0 = [[x, 0] for x in negative] 115 | gt1 = [[x, 1] for x in positive] 116 | gt = gt0+gt1 117 | if len(gt) == num_dt: 118 | pass 119 | elif len(gt) != num_dt and 2*len(gt) == num_dt: 120 | gt_fogged = [[f'fogged_{x}', cls] for x, cls in gt] 121 | gt += gt_fogged 122 | else: 123 | raise NameError(f'num_gt: {len(gt)} is not compatible with num_dt: {num_dt}') 124 | 125 | gt_data = np.array(gt) 126 | return gt_data 127 | 128 | def get_detect_data(self, v5_out=None): 129 | """ 130 | :param v5_out: 131 | :return: dt_data, an array, shape(nfiles, 50, 2) 132 | nfiles: number of output files, 100 represents a maximum of 100 targets in an image. 133 | 2 represents class and confidence. 134 | dt_data[dt_data[:, :, 1] != -2]，筛选有目标的jpg,filename会重复。shape, (n_detected_files, 3) 检测到的目标的图片数目。 135 | """ 136 | files = os.listdir(v5_out) 137 | imgs = sorted([tmp.strip('\n') for tmp in files if tmp.endswith(self.img_suffix)]) 138 | txts = [tmp.strip('\n') for tmp in files if tmp.endswith('.txt')] 139 | 140 | dt_data = np.zeros((len(imgs), 100, 7), dtype=np.object_) # 7: img cls conf bbox(xcycwh_percent) 141 | # 最多100个目标 142 | dt_data[:, :, :] = -2 143 | cls_c = np.zeros(100).astype(np.int32) # 分类count 144 | for i, img in enumerate(imgs): 145 | file_stem = Path(img).stem 146 | txt_file = f'{v5_out}/{file_stem}.txt' 147 | if os.path.exists(txt_file): 148 | with open(txt_file, 'r') as f: 149 | txt_data = f.readlines() 150 | # if len(txt_data) > 1: 151 | # print(img) 152 | for j, t in enumerate(txt_data): # 153 | cls, conf, xc, yc, w, h = t.split() 154 | # tmp = [img, int(cls), float(conf)] 155 | tmp = [img, int(cls), float(conf), float(xc), float(yc), float(w), float(h)] 156 | dt_data[i, j, :] = np.array(tmp) 157 | cls_c[int(cls)] += 1 158 | else: 159 | dt_data[i, 0, :] = np.array([img, -2, -2, -2, -2, -2, -2]) 160 | 161 | cls_c = cls_c[cls_c[:] != 0] 162 | print( 163 | f'detected output: {len(imgs)} imgs[.{self.img_suffix}], {len(txts)} txts. {np.sum(cls_c)} detected targets.') 164 | prnts = '\n'.join([f'class{i}: {count} targets detected' for i, count in enumerate(cls_c)]) 165 | print(prnts) 166 | return dt_data 167 | 168 | def detect(self): 169 | self.net.eval() 170 | os.system(r'rm -rf detect_ret/detect.txt') 171 | for i, filename in enumerate(self.cfg.filenames): 172 | t = time.time() 173 | img_path = f"{self.cfg.infer_img_dir}/{filename}" 174 | img = Image.open(img_path) 175 | img, outs = self.infer(img) 176 | # img.save(f'/Users/tanmenglu/Downloads/AutoDrive/ret/{filename}') 177 | print(f"{filename} time: {time.time()-t:.2f} FPS: {1/(time.time()-t):.2f}") 178 | strs = '' 179 | for ot in outs: 180 | strs += ' ' + ','.join(ot) 181 | os.system(f'echo "{filename}{strs}" >>detect_ret/detect.txt') 182 | 183 | 184 | def plot(): 185 | import matplotlib.pyplot as plt 186 | threshes = [i / 100 for i in range(10, 91, 2)] 187 | xs = [] 188 | ys = [] 189 | for thresh in threshes: 190 | ret = GCM(thresh=thresh, v5_out=v5_out) 191 | print(ret) 192 | ret = ret.split('\n')[-2].split() 193 | xs.append((ret[1])) 194 | ys.append([ret[3].strip('%'), ret[5].strip('%'), ret[7].strip('%'), ret[9].strip('%'), ret[11].strip('%')]) 195 | 196 | xs = np.array(xs).astype(np.float).reshape((-1, 1)) 197 | xs = np.repeat(xs, 5, axis=1) 198 | ys = np.array(ys).astype(np.float) 199 | print(xs.shape, ys.shape) 200 | print(xs, ys) 201 | lbs = ['ACC', 'PPV', 'RECALL', 'TNR', 'F1_score'] 202 | colors = [[np.random.randint(0, 255) for _ in range(3)] for _ in range(xs.shape[1])] 203 | colors = [[tmp/255 for tmp in tmp2] for tmp2 in colors] 204 | for i in range(xs.shape[1]): 205 | x = xs[:, i] 206 | y = ys[:, i] 207 | print(colors[i]) 208 | plt.plot(x, y, label=lbs[i], color=colors[i]) 209 | 210 | # 标出max： 211 | ys_max = np.max(ys, axis=0) # 每一列是一个种类，对每一列求最大值m 212 | ys_max_i = np.argmax(ys, axis=0) # [c1,c2,c3,c4,c5] 213 | xs_max = xs[:, 0][ys_max_i] 214 | print(xs_max, ys_max_i, ys_max) 215 | for i in range(len(xs_max)): 216 | x = xs_max[i] 217 | y = ys_max[i] 218 | plt.text(x, y, f"{lbs[i]} {x} {y}", color=colors[i]) 219 | plt.plot(x, y, 'o', color=colors[i]) 220 | 221 | plt.xticks() 222 | plt.legend(loc='best') 223 | plt.xlabel('thresh') 224 | plt.ylabel('precent/%') 225 | plt.savefig('xxx.png', bbox_inches='tight') 226 | 227 | 228 | if __name__ == "__main__": 229 | GCM = GetConfusingMatrix() 230 | 231 | path = "/home/zzd/datasets/crosswalk/testsets_1770" 232 | ret_all = [] 233 | head = f"trs tes {'ALL':<5}{'TP':<5}{'FP':<5}{'FN':<5}{'TN':<5} thr acc P R TNR F1 {'time':<7} fps\n" 234 | ret_all.append(head) 235 | train_sizes = [640] 236 | test_sizes = range(128, 160, 32) 237 | 238 | exps = [ 239 | 'm_ep300', 'SEm_NST0_fog0_ep100', 'SEm_NST1_fog0_ep100', 240 | 'SEm_NST1_fog0_ep300', 'SEm_NST1_fog1_ep100'] 241 | 242 | for trs in train_sizes: # train_size 243 | for tes in test_sizes: # test_size 244 | v5_out = f"{path}/dense_ts_roi_SE300epoch_{trs}_{tes}_output" 245 | v5_out = f'{path}/SEv5m_300EP_no_roi' 246 | # v5_out = f'{path}/SEYOLOv5m_ep100_exp' 247 | # v5_out = f'{path}/m_ep300_sz640_ROI1_SSVM1' 248 | v5_out = f'{path}/SEm_NST0_fog0_ep100_sz640_ROI1_SSVM1' 249 | 250 | # v5_out = f'{path}/SEYOLOv5m_ep100_exp_sz640_roi1_ssvm1' 251 | # v5_out = f'{path}/SEYOLOv5m_ep100_exp_sz640_roi1_ssvm1' 252 | ret = GCM.analyse(thresh=0.5, v5_out=v5_out, pn_dir=path) 253 | # 读取inference时间 254 | with open(f'{v5_out}/detect.log', 'r') as f: 255 | log = f.readlines() 256 | time, fps = re.findall(f'\d+.\d+', log[-1]) 257 | time = time[:-1] 258 | fps = fps[:-3] 259 | ret_all.append(f'{trs} {tes} {ret} {time} {fps}\n') 260 | with open('confusing_matrix_dense_roi_SE300epoch.txt', 'w') as f: 261 | f.writelines(ret_all) 262 | -------------------------------------------------------------------------------- /scripts/post.bak.py: -------------------------------------------------------------------------------- 1 | """ 2 | post for crosswalk detection, the vector crossing method used to detect the vehicle crossing crosswalk behavior 3 | 这个bak是添加抖动保护前的备份。20200929 4 | """ 5 | 6 | import os 7 | import cv2 8 | import numpy as np 9 | import yaml 10 | 11 | __author__ = 'zhangzhengde@mail.sjtu.edu.cn' 12 | __date__ = '20200929' 13 | 14 | 15 | class DmPost(object): 16 | """damei post""" 17 | def __init__(self, opt): 18 | self.opt = opt 19 | self.crossout = self.init_crossout() 20 | self.vetors = self.init_vectors() 21 | self.logo = self.init_logo() 22 | self.rect = self.init_rect() 23 | self.t_inter = self.init_others() 24 | self.cl, self.rl = self.init_clrl() 25 | 26 | def init_crossout(self): 27 | crossout = np.zeros( 28 | (self.opt.field_size, 8)) 29 | # 5 rows. 8 columns ：out_index, detect_id, fileorder, is crosswalk exists, xc, yc, count，recording_flag 30 | crossout[:, :] = -2 # -2 represents undetected 31 | crossout[:, 0] = range(len(crossout)) # init out_index 32 | crossout[:, 6:8] = 0 # init count and recording_flag to 0 33 | return crossout 34 | 35 | def init_vectors(self): 36 | vector_size = 600 * 30 # maximun 600 seconds, sampling ratio 30。 37 | vectors = np.zeros((vector_size, 2)) # 2: store xc, yc 38 | vectors[:, :] = -2 # init 39 | return vectors 40 | 41 | def init_logo(self): 42 | logo_path = "others/SHNAVI.jpg" 43 | logo = cv2.imread(logo_path) 44 | imgs = [x for x in os.listdir(self.opt.source) if x.endswith('jpg') or x.endswith('png')] 45 | if len(imgs) == 0: 46 | raise NameError(f'did not get imgs from {self.opt.source}, support suffix: .jpg .png') 47 | img_size = cv2.imread(f'{self.opt.source}/{imgs[0]}').shape[:2] 48 | logo_size = logo.shape[:2] # 248, 1217 49 | scale_factor = 3.5 50 | resized_logo_size = ( 51 | int(img_size[1] / scale_factor), int(img_size[1] / scale_factor * logo_size[0] / logo_size[1])) 52 | logo = cv2.resize(logo, resized_logo_size, interpolation=cv2.INTER_LINEAR) 53 | return logo 54 | 55 | def init_rect(self): 56 | rect = cv2.imread("others/rect.jpg") 57 | rect_resized_size = (self.logo.shape[1], 200) 58 | rect = cv2.resize(rect, rect_resized_size, interpolation=cv2.INTER_LINEAR) 59 | return rect 60 | 61 | def init_clrl(self): 62 | # get opt 63 | opt = self.opt 64 | with open(opt.control_line_setting, 'r') as f: 65 | conls = yaml.load(f, Loader=yaml.FullLoader) 66 | print(conls) 67 | cl = conls[opt.select_control_line] # control line 68 | rl = conls['red_line'] # red_line 69 | return cl, rl 70 | 71 | def init_others(self): 72 | FPS = 30 # 1s 有25张图 73 | sampling_ratio = 30 # 每4张采样1张 74 | sampling_rate = FPS / sampling_ratio # 采样率,1秒5张 75 | base_time = 0 # 基准时间，检测的图片超过chunk大小时，基准时间要增加 76 | t_inter = 1 / sampling_rate # time interpolation 每张0.2秒 77 | return t_inter 78 | 79 | def dmpost(self, img, det, det_id, filename, names, cls='crosswalk'): 80 | """ 81 | :param img: 82 | :param det: None或tensor，shape:(n, 6), n是n个目标，6是已经scale到原图的x1,y1,x2,y2和confidence、class 83 | :param filename: filename0001.jpg获取order，或者输入0036_1，0036是视频名，1是第1帧。 84 | :param names: 模型中的类名 85 | :param cls: 后处理筛选的分类，默认是crosswalk 86 | :return: 87 | """ 88 | opt = self.opt 89 | crossout = self.crossout 90 | vectors = self.vetors 91 | 92 | # 0000 0001 0002 0003 and so on # 1 2 3 4 .... 93 | fileorder = int(filename.split('.')[0][-4::]) if filename.endswith('jpg') else int(filename.split('_')[-1]) 94 | 95 | # 绘制上下控制线 96 | # cl = opt.control_line # control line, [360, 700] 97 | cl = self.cl 98 | rl = self.rl 99 | lt = round(0.002 * (img.shape[0] + img.shape[1]) / 2) + 1 # line thickness 100 | tf = max(lt - 1, 1) # 2 101 | h, w = img.shape[:2] 102 | top = [(0, cl[0]), (w, cl[0])] 103 | button = [(0, cl[1]), (w, cl[1])] 104 | middle = [(int(w / 4), int(cl[0]+(cl[1]-cl[0])*rl)), (int(3 * w / 4), int(cl[0]+(cl[1]-cl[0])*rl))] 105 | arrow = [(int(w / 2), middle[0][1]), (int(w / 2), middle[0][1] + 50)] 106 | textpos = (top[0][0] + 10, top[0][1] + 30) 107 | 108 | cv2.line(img, top[0], top[1], color=(30, 30, 224), thickness=lt) 109 | cv2.line(img, button[0], button[1], color=(30, 30, 224), thickness=lt) 110 | cv2.line(img, middle[0], middle[1], color=(120, 85, 220), thickness=lt) # 粉色 111 | cv2.arrowedLine(img, arrow[0], arrow[1], (120, 85, 220), 5, 8, 0, 0.3) 112 | cv2.putText(img, 'Control Line', textpos, 0, lt / 3, (30, 30, 224), thickness=tf, lineType=cv2.LINE_AA) 113 | 114 | cls_idx = names.index(cls) # class index i.e. 0 for crosswalk 115 | det_include_class = cls_idx in det[:, 5] if det is not None else False 116 | if det is not None and det_include_class: 117 | det = det.cpu().numpy() 118 | crosswalk = np.zeros((det[det[:, 5] == 0].shape[0], 8)) 119 | # [n, 6]变成[n, 8]，去掉cls。8分别是x1,y1,x2,y2,conf,cx,cy,is_in_control_line 120 | # print(det.shape, det[:, 5] == 1) 121 | crosswalk[:, :5] = det[det[:, 5] == 0][:, :5] # 筛选只要crosswalk，默认起序号为0, 0 1 2 3 4 122 | cx = (crosswalk[:, 0] + crosswalk[:, 2]) // 2 123 | cy = (crosswalk[:, 1] + crosswalk[:, 3]) // 2 124 | crosswalk[:, 5] = cx 125 | crosswalk[:, 6] = cy 126 | is_in_cl = (crosswalk[:, 6] > cl[0]) | (crosswalk[:, 6] < cl[1]) 127 | crosswalk[:, 7] = is_in_cl 128 | if crosswalk.shape[0] > 1: 129 | # 同时检测到多条斑马线，根据之前记录的crossout，确定选择哪一条 130 | # lastxc, lastyc = crossout 131 | last_co = crossout[crossout[:, 3] != -2] # last crossout 132 | # print(last_co, last_co.shape) 133 | if len(last_co) == 0: # 都是空的，，那就使用第一条 134 | valid_idx = 0 135 | else: # 计算检测到的斑马线的中心与上次记录的中心哪个接近就用哪个 136 | lastcxy = last_co[-1][4:6] 137 | currentcxy = crosswalk[:, 5:7] 138 | # print(lastcxy, lastcxy.shape, currentcxy, currentcxy.shape) 139 | distances = np.sum(np.square(currentcxy - lastcxy), axis=1) # 距离 140 | valid_idx = np.argmin(distances) 141 | print(f'WANING: detect.py post, the detected crosswalk is more than one, use the {valid_idx+1} one') 142 | crosswalk = crosswalk[valid_idx, :].reshape(1, -1) 143 | det = det[valid_idx, :].reshape(1, -1) 144 | else: 145 | crosswalk = np.zeros((1, 8)) 146 | 147 | # print(crosswalk.shape) 148 | # print(det.shape) 149 | 150 | if det_id < crossout.shape[0]: # detected_img id < 5 151 | # 该列更新值 n列：out_index, detect_id, fileorder, 有无斑马线, xc, yc 152 | crossout[det_id, 1] = det_id 153 | crossout[det_id, 2] = fileorder 154 | crossout[det_id, 3] = crosswalk[0, 7] 155 | crossout[det_id, 4:6] = crosswalk[0, 5: 7] # xc, yc 156 | index = det_id 157 | else: 158 | crossout[0:-1:, 1::] = crossout[1::, 1::] # 除了序号列的所有行向上平移一格 159 | # 最后一列更新值 160 | crossout[-1, 1] = det_id 161 | crossout[-1, 2] = fileorder 162 | crossout[-1, 3] = crosswalk[0, 7] 163 | crossout[-1, 4:6] = crosswalk[0, 5: 7] # xc, yc 164 | index = len(crossout) - 1 165 | 166 | # print(crossout[:11, :]) 167 | exist, vector, scale = self.decode_crossout(crossout, index) 168 | recording = crossout[index, 7] 169 | 170 | if recording == 1 and vector is not None: 171 | vectors[opt.v_idx, :] = vector[0] 172 | vectors[opt.v_idx + 1, :] = vector[1] 173 | opt.v_idx += 1 174 | elif recording == 1 and vector is None: 175 | pass 176 | else: 177 | vectors[:, :] = -2 178 | opt.v_idx = 0 179 | 180 | speed = None if scale is None else float((vector[1][1] - vector[0][1]) / (self.t_inter * scale)) 181 | # 向量的y距离除以scale再除以time interpolate 182 | 183 | # 绘制logo和结果 184 | # logo_pos = (int(img.shape[1]/2-logo.shape[1]/2), cl[1]-logo.shape[0]) 185 | logo_pos = (20, 20) # w h 186 | img = self.imgAdd(self.logo, img, logo_pos[1], logo_pos[0], alpha=0.5) 187 | rect_pos = (20, 20 + 5 + self.logo.shape[0]) 188 | img = self.imgAdd(self.rect, img, rect_pos[1], rect_pos[0], alpha=0.5) 189 | 190 | pos = (20 + 20, 20 + 5 + self.logo.shape[0] + 20 + 30) 191 | self.imgputText(img, f'crosswalk: {exist}', pos, lt, tf) 192 | pos = (20 + 20, 20 + 5 + self.logo.shape[0] + 20 + 30 + 40) 193 | self.imgputText(img, f'speed: {speed:.2f}', pos, lt, tf) if speed is not None else None 194 | self.imgputText(img, f'speed: {speed}', pos, lt, tf) if speed is None else None 195 | 196 | vt = vectors[vectors[:, 0] != -2].astype(np.int) # 筛选t 197 | self.imgplotVectors(img, vt) if vector is not None else None 198 | 199 | # 计数，vectors有值、向量经过中间线、当前vector不为None时，count+1 200 | crossout[index, 6] = crossout[index - 1, 6] # 先同步count与前一个相同 201 | if vt.shape[0] != 0: 202 | # print(vt.shape) 203 | if vt[0, 1] < np.mean(cl): 204 | intersect = vt[vt[:, 1] > middle[0][1]] 205 | # print('inter', intersect) 206 | if intersect.shape[0] == 1 and vector is not None: 207 | crossout[index, 6] += 1 # count+1 208 | prt_str = f'\n{filename} count+1, conf: {crosswalk[0, 4]:.2f} count: {int(crossout[index, 6])}' 209 | print(prt_str) 210 | os.system(f'echo "{prt_str}" >> {opt.output}/detect.log') 211 | count = int(crossout[index, 6]) 212 | 213 | pos = (20 + 20, 20 + 5 + self.logo.shape[0] + 20 + 30 + 40 + 40) 214 | self.imgputText(img, f'count: {count}', pos, lt, tf) 215 | 216 | # 打印状态：有vectors时就是crossing 217 | pos = (20 + 20, 20 + 5 + self.logo.shape[0] + 20 + 30 + 40 + 40 + 40) 218 | status = 'No crosswalk' if vt.shape[0] == 0 else 'Crossing' 219 | self.imgputText(img, f'status: {status}', pos, lt, tf) 220 | 221 | prt_str = f'{filename} detect_id: {det_id} speed: {speed} count: {count} status: {status}' 222 | print(prt_str) 223 | os.system(f'echo "{prt_str}" >> {opt.output}/detect.log') 224 | 225 | return img 226 | 227 | def decode_crossout(self, crossout, index, vector_threash=20, vector_max_threash=600): 228 | """ 229 | 解码crossout, 输出当前图是否有斑马线，斑马线位移矢量，时间尺度（索引的间距） 230 | count计数算法：该矢量经过中心线时，计数+1。 231 | :param crossout: 232 | :index: 当前位置索引 233 | :recursive_bits: 向前递归位数 234 | :vector_threash: 向量的模的阈值，大于该阈值且y的变化为负数时，设置recording为0，断开count。 235 | :return: 236 | """ 237 | exist = crossout[index, 3] 238 | co = crossout[crossout[:, 3] == 1] # 所有存在的行 239 | if exist == 0: 240 | if co.shape[0] == 0: # 感受野范围内都没有斑马线 241 | crossout[:, 7] = 0 # recording置0 242 | return False, None, None 243 | else: 244 | if co.shape[0] == 1: # 只有最后一行有 245 | crossout[:, 7] = 0 # recording置0 246 | return False, None, None 247 | else: 248 | scale = co[-1, 1] - co[-2, 1] # detected_id的差 249 | vector = [co[-2, 4:6], co[-1, 4:6]] # 2个点 ((xc1, yc1), (xc2, yc2)) 250 | vector2 = vector[1] - vector[0] # (x2-x1) (y2-y1) 251 | length = np.sqrt(vector2[0] ** 2 + vector2[1] ** 2) 252 | y_shift = vector2[1] 253 | # print(vector2, vector2/1280, vector2/720, length, y_shift) 254 | if length > vector_threash and y_shift < 0: 255 | crossout[:, 7] = 0 # recording置0 256 | elif length > vector_max_threash: # 有时候会出现超级长的大于300像素，筛掉，约((680-400)*2/3)**2 680和400是控制线 257 | crossout[:, 7] = 0 258 | else: 259 | crossout[:, 7] = 1 # recording置1 260 | # print(crossout[:, 7], length) 261 | return True, vector, scale 262 | 263 | def imgAdd(self, small_img, big_image, x, y, alpha=0.5): 264 | """ 265 | 把小图贴到大图的xy位置，透明度设置为0.5 266 | """ 267 | row, col = small_img.shape[:2] 268 | if small_img.shape[0] > big_image.shape[0] or small_img.shape[1] > big_image.shape[1]: 269 | raise NameError(f'imgAdd, the size of small img bigger than big img.') 270 | roi = big_image[x:x + row, y:y + col, :] 271 | roi = cv2.addWeighted(small_img, alpha, roi, 1 - alpha, 0) 272 | big_image[x:x + row, y:y + col] = roi 273 | return big_image 274 | 275 | def imgputText(self, img, txt, pos, lt, tf): 276 | cv2.putText(img, txt, pos, 0, lt / 3, (30, 30, 224), thickness=tf, lineType=cv2.LINE_AA) 277 | 278 | def imgplotVectors(self, img, vt): 279 | if vt.shape[0] == 0: 280 | return 281 | for i in range(vt.shape[0] - 1): 282 | pt1, pt2 = tuple(vt[i]), tuple(vt[i + 1]) 283 | cv2.arrowedLine(img, pt1, pt2, (0, 0, 255), 5, 8, 0, 0.3) 284 | -------------------------------------------------------------------------------- /scripts/pt2wts.py: -------------------------------------------------------------------------------- 1 | """ 2 | convert xxx.pt to xxx.wts 3 | 4 | need install damei library: 5 | pip install damei -i https://pypi.Python.org/simple 6 | """ 7 | import damei as dm 8 | import argparse 9 | 10 | parser = argparse.ArgumentParser() 11 | parser.add_argument( 12 | 'weights', type=str, default='runs/exp0/weights/best.pt', 13 | help='path to weights [xx.pt]') 14 | parser.add_argument('--output_dir', type=str, default=None) 15 | opt = parser.parse_args() 16 | 17 | dm.post.pt2wts(opt.weights, opt.output_dir) 18 | 19 | -------------------------------------------------------------------------------- /scripts/synthetic_fog.py: -------------------------------------------------------------------------------- 1 | """ 2 | 直接运行程序可以测试合成雾气效果 3 | Produced by: zhangzhengde@sjtu.edu.cn 4 | """ 5 | import os 6 | import math 7 | import cv2 8 | import time 9 | from pathlib import Path 10 | import numpy as np 11 | import matplotlib.pyplot as plt 12 | from tqdm import tqdm 13 | import shutil 14 | 15 | 16 | class SyntheticFog(object): 17 | def __init__(self): 18 | pass 19 | 20 | def __call__(self, show=False): 21 | img_path = '../example/fog_image/raw.jpg' 22 | # img_path = '../sources/IMG_6685.JPG' 23 | assert os.path.exists(img_path), f'error: img does not exists, {img_path}' 24 | img = cv2.imread(img_path) 25 | print(img.shape) 26 | img = img/255.0 27 | print(f'fogging...') 28 | t0 = time.time() 29 | br = 0.7 30 | th = 0.05 31 | fogged_img = self.fogging_img( 32 | img, brightness=br, thickness=th, 33 | high_efficiency=True) 34 | print(f'fogging time: {(time.time()-t0)*1000:.4f}ms') 35 | rf = 1 # resize factor 36 | img = cv2.resize(img, (int(img.shape[1]*rf), int(img.shape[0]*rf))) 37 | fogged_img = cv2.resize(fogged_img, ((int(fogged_img.shape[1]*rf)), (int(fogged_img.shape[0]*rf)))) 38 | fogged_img = np.array(fogged_img*255, dtype=np.uint8) 39 | if show: 40 | cv2.imshow('src', img) 41 | cv2.imshow('fogged', fogged_img) 42 | cv2.waitKey(0) 43 | cv2.imwrite(f'../example/fog_image/fogged_br{br}_th{th}.jpg', fogged_img) 44 | 45 | def fogging_dir(self, sp, tp=None, random_params=True, brightness=None, thickness=None, save_src_img=False): 46 | """ 47 | fogging images in a directory 48 | :param sp: str, source dir path 49 | :param tp: str, target dir path, tp is fogged_{sp} by default 50 | :param random_params: bool, use random brightness and fog thickness params if True 51 | :param brightness: float, 0.1 to 0.9, gray of synthetic fog, pure white fog if 1, dark fog if 0. 52 | :param thickness: float, 0.01 to 0.09, thickness of synthetic fog, the larger the value, the thicker the fog. 53 | :param save_src_img: save source image at the same time i.e. copy source imgs to tp 54 | :return: None, all fogged images will be saved to target dir path. 55 | """ 56 | tp = tp if tp is not None else f'{Path(sp).parent}/fogged_{Path(sp).name}' 57 | if os.path.exists(tp): 58 | ipt = input(f'Target dir: {tp} exists, do you want to remove it and continue. [Yes]/No: ') 59 | if ipt in ['', 'Yes', 'Y', 'yes']: 60 | shutil.rmtree(tp) 61 | else: 62 | print('do nothing') 63 | exit() 64 | os.makedirs(f'{tp}') 65 | 66 | imgs = [x for x in os.listdir(sp) if str(Path(x).suffix).lower() in ['.jpg', '.bmp']] 67 | print(f'Fogging {len(imgs)} images in dir {sp}, \nfogged images will be save to {tp}.') 68 | bar = tqdm(imgs) 69 | for i, img_name in enumerate(bar): 70 | img_path = f'{sp}/{img_name}' 71 | # stem = Path(img_path).stem 72 | # suffix = Path(img_path).suffix 73 | 74 | if save_src_img: # save source img 75 | shutil.copy(f'{sp}/{img_name}', f'{tp}/{img_name}') 76 | 77 | img = cv2.imread(img_path) 78 | h, w, c = img.shape 79 | normed_img = img.copy()/255.0 80 | 81 | if random_params: 82 | br = np.clip(0.2 * np.random.randn() + 0.5, 0.1, 0.9) # 0.1~0.9 83 | th = np.clip(0.01 * np.random.randn() + 0.05, 0.01, 0.09) 84 | else: 85 | br = brightness 86 | th = thickness 87 | assert br is not None 88 | assert th is not None 89 | fogged_img = self.fogging_img(normed_img, br, th, high_efficiency=True) 90 | fogged_img = np.array(fogged_img * 255, dtype=np.uint8) 91 | cv2.imwrite(f'{tp}/fogged_{img_name}', fogged_img) 92 | 93 | bar.set_description(f'Fogged image saved, fogged_{img_name}') 94 | 95 | def fogging_img(self, img, brightness=0.7, thickness=0.05, high_efficiency=True): 96 | """ 97 | fogging single image 98 | :param img: src img 99 | :param brightness: brightness 100 | :param thickness: fog thickness, without fog when 0, max 0.1, 101 | :param high_efficiency: use matrix to improve fogging speed when high_efficiency is True, else use loops 102 | low efficiency: about 4000ms, high efficiency: about 80ms, tested in (864, 1152, 3) img 103 | :return: fogged image 104 | """ 105 | assert 0 <= brightness <= 1 106 | assert 0 <= thickness <= 0.1 107 | fogged_img = img.copy() 108 | h, w, c = fogged_img.shape 109 | if not high_efficiency: # use default loop to fogging, low efficiency 110 | size = np.sqrt(np.max(fogged_img.shape[:2])) # 雾化尺寸 111 | center = (h // 2, w // 2) # 雾化中心 112 | # print(f'shape: {img.shape} center: {center} size: {size}') # 33 113 | # d_list = [] 114 | for j in range(h): 115 | for l in range(w): 116 | d = -0.04 * math.sqrt((j - center[0]) ** 2 + (l - center[1]) ** 2) + size 117 | # print(f'd {d}') 118 | td = math.exp(-thickness * d) 119 | # d_list.append(td) 120 | fogged_img[j][l][:] = fogged_img[j][l][:] * td + brightness * (1 - td) 121 | # x = np.arange(len(d_list)) 122 | # plt.plot(x, d_list, 'o') 123 | # if j == 5: 124 | # break 125 | else: # use matrix # TODO: 直接使用像素坐标，距离参数不适用于大分辨率图像，会变成鱼眼镜头的样子. done. 126 | use_pixel = True 127 | size = np.sqrt(np.max(fogged_img.shape[:2])) if use_pixel else 1 # 雾化尺寸 128 | h, w, c = fogged_img.shape 129 | hc, wc = h // 2, w // 2 130 | mask = self.get_mask(h=h, w=w, hc=hc, wc=wc, pixel=use_pixel) # (h, w, 2) 131 | d = -0.04 * np.linalg.norm(mask, axis=2) + size 132 | 133 | td = np.exp(-thickness * d) 134 | 135 | for cc in range(c): 136 | fogged_img[..., cc] = fogged_img[..., cc] * td + brightness*(1-td) 137 | 138 | # a = np.linalg.norm(mask, axis=2) 139 | # print(f'size: {fogged_img.shape} a: {a} max: {np.max(fogged_img)} {np.min(fogged_img)}') 140 | 141 | fogged_img = np.clip(fogged_img, 0, 1) # 解决黑白噪点的问题 142 | # print(f'mask: {mask[:, :, 1]} {mask.shape}') 143 | # print(f'd: {d} {d.shape}') 144 | 145 | return fogged_img 146 | 147 | def get_mask(self, h, w, hc, wc, pixel=True): 148 | mask = np.zeros((h, w, 2), dtype=np.float32) 149 | if pixel: 150 | mask[:, :, 0] = np.repeat(np.arange(h).reshape((h, 1)), w, axis=1) - hc 151 | mask[:, :, 1] = np.repeat(np.arange(w).reshape((1, w)), h, axis=0) - wc 152 | else: 153 | mask[:, :, 0] = np.repeat(np.linspace(0, 1, h).reshape(h, 1), w, axis=1) - 0.5 154 | mask[:, :, 1] = np.repeat(np.linspace(0, 1, w).reshape((1, w)), h, axis=0) - 0.5 155 | return mask 156 | 157 | 158 | if __name__ == '__main__': 159 | synf = SyntheticFog() 160 | synf(show=True) 161 | 162 | 163 | 164 | -------------------------------------------------------------------------------- /scripts/yolo2coco.py: -------------------------------------------------------------------------------- 1 | """ 2 | 转换训练集，YOLOv5 format to COCO format 3 | """ 4 | import os, sys 5 | sys.path.append(f'/home/zzd/PycharmProject/damei') 6 | import damei as dm 7 | 8 | sp = '/home/zzd/datasets/crosswalk/fogged_train_data_v5_format' 9 | tp = "/home/zzd/datasets/crosswalk/fogged_train_data_coco_format" 10 | dm.tools.yolo2coco(sp=sp, tp=tp) 11 | 12 | -------------------------------------------------------------------------------- /utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zhangzhengde0225/CDNet/6ffc9a6a28398ecea78c05bd38cc37dc4ce6a252/utils/__init__.py -------------------------------------------------------------------------------- /utils/google_utils.py: -------------------------------------------------------------------------------- 1 | # This file contains google utils: https://cloud.google.com/storage/docs/reference/libraries 2 | # pip install --upgrade google-cloud-storage 3 | # from google.cloud import storage 4 | 5 | import os 6 | import time 7 | from pathlib import Path 8 | 9 | 10 | def attempt_download(weights): 11 | # Attempt to download pretrained weights if not found locally 12 | weights = weights.strip().replace("'", '') 13 | msg = weights + ' missing, try downloading from https://github.com/zhangzhengde0225/CDNet/blob/master/docs/MODEL_ZOO.md' 14 | 15 | r = 1 # return 16 | if len(weights) > 0 and not os.path.isfile(weights): 17 | d = {'yolov3-spp.pt': '1mM67oNw4fZoIOL1c8M3hHmj66d8e-ni_', # yolov3-spp.yaml 18 | 'yolov5s.pt': '1R5T6rIyy3lLwgFXNms8whc-387H0tMQO', # yolov5s.yaml 19 | 'yolov5m.pt': '1vobuEExpWQVpXExsJ2w-Mbf3HJjWkQJr', # yolov5m.yaml 20 | 'yolov5l.pt': '1hrlqD1Wdei7UT4OgT785BEk1JwnSvNEV', # yolov5l.yaml 21 | 'yolov5x.pt': '1mM8aZJlWTxOg7BZJvNUMrTnA2AbeCVzS', # yolov5x.yaml 22 | } 23 | 24 | file = Path(weights).name 25 | if file in d: 26 | r = gdrive_download(id=d[file], name=weights) 27 | 28 | if not (r == 0 and os.path.exists(weights) and os.path.getsize(weights) > 1E6): # weights exist and > 1MB 29 | os.remove(weights) if os.path.exists(weights) else None # remove partial downloads 30 | s = "curl -L -o %s 'storage.googleapis.com/ultralytics/yolov5/ckpt/%s'" % (weights, file) 31 | r = os.system(s) # execute, capture return values 32 | 33 | # Error check 34 | if not (r == 0 and os.path.exists(weights) and os.path.getsize(weights) > 1E6): # weights exist and > 1MB 35 | os.remove(weights) if os.path.exists(weights) else None # remove partial downloads 36 | raise Exception(msg) 37 | 38 | 39 | def gdrive_download(id='1n_oKgR81BJtqk75b00eAjdv03qVCQn2f', name='coco128.zip'): 40 | # Downloads a file from Google Drive, accepting presented query 41 | # from utils.google_utils import *; gdrive_download() 42 | t = time.time() 43 | 44 | print('Downloading https://drive.google.com/uc?export=download&id=%s as %s... ' % (id, name), end='') 45 | os.remove(name) if os.path.exists(name) else None # remove existing 46 | os.remove('cookie') if os.path.exists('cookie') else None 47 | 48 | # Attempt file download 49 | os.system("curl -c ./cookie -s -L \"drive.google.com/uc?export=download&id=%s\" > /dev/null" % id) 50 | if os.path.exists('cookie'): # large file 51 | s = "curl -Lb ./cookie \"drive.google.com/uc?export=download&confirm=`awk '/download/ {print $NF}' ./cookie`&id=%s\" -o %s" % ( 52 | id, name) 53 | else: # small file 54 | s = 'curl -s -L -o %s "drive.google.com/uc?export=download&id=%s"' % (name, id) 55 | r = os.system(s) # execute, capture return values 56 | os.remove('cookie') if os.path.exists('cookie') else None 57 | 58 | # Error check 59 | if r != 0: 60 | os.remove(name) if os.path.exists(name) else None # remove partial 61 | print('Download error ') # raise Exception('Download error') 62 | return r 63 | 64 | # Unzip if archive 65 | if name.endswith('.zip'): 66 | print('unzipping... ', end='') 67 | os.system('unzip -q %s' % name) # unzip 68 | os.remove(name) # remove zip to free space 69 | 70 | print('Done (%.1fs)' % (time.time() - t)) 71 | return r 72 | 73 | 74 | # def upload_blob(bucket_name, source_file_name, destination_blob_name): 75 | # # Uploads a file to a bucket 76 | # # https://cloud.google.com/storage/docs/uploading-objects#storage-upload-object-python 77 | # 78 | # storage_client = storage.Client() 79 | # bucket = storage_client.get_bucket(bucket_name) 80 | # blob = bucket.blob(destination_blob_name) 81 | # 82 | # blob.upload_from_filename(source_file_name) 83 | # 84 | # print('File {} uploaded to {}.'.format( 85 | # source_file_name, 86 | # destination_blob_name)) 87 | # 88 | # 89 | # def download_blob(bucket_name, source_blob_name, destination_file_name): 90 | # # Uploads a blob from a bucket 91 | # storage_client = storage.Client() 92 | # bucket = storage_client.get_bucket(bucket_name) 93 | # blob = bucket.blob(source_blob_name) 94 | # 95 | # blob.download_to_filename(destination_file_name) 96 | # 97 | # print('Blob {} downloaded to {}.'.format( 98 | # source_blob_name, 99 | # destination_file_name)) 100 | -------------------------------------------------------------------------------- /utils/torch_utils.py: -------------------------------------------------------------------------------- 1 | import math 2 | import os 3 | import time 4 | from copy import deepcopy 5 | 6 | import torch 7 | import torch.backends.cudnn as cudnn 8 | import torch.nn as nn 9 | import torch.nn.functional as F 10 | import torchvision.models as models 11 | 12 | 13 | def init_seeds(seed=0): 14 | torch.manual_seed(seed) 15 | 16 | # Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html 17 | if seed == 0: # slower, more reproducible 18 | cudnn.deterministic = True 19 | cudnn.benchmark = False 20 | else: # faster, less reproducible 21 | cudnn.deterministic = False 22 | cudnn.benchmark = True 23 | 24 | 25 | def select_device(device='', batch_size=None): 26 | # device = 'cpu' or '0' or '0,1,2,3' 27 | cpu_request = device.lower() == 'cpu' 28 | if device and not cpu_request: # if device requested other than 'cpu' 29 | os.environ['CUDA_VISIBLE_DEVICES'] = device # set environment variable 30 | assert torch.cuda.is_available(), 'CUDA unavailable, invalid device %s requested' % device # check availablity 31 | 32 | cuda = False if cpu_request else torch.cuda.is_available() 33 | if cuda: 34 | c = 1024 ** 2 # bytes to MB 35 | ng = torch.cuda.device_count() 36 | # zzd k80 37 | # ng = 2 38 | if ng > 1 and batch_size: # check that batch_size is compatible with device_count 39 | assert batch_size % ng == 0, 'batch-size %g not multiple of GPU count %g' % (batch_size, ng) 40 | x = [torch.cuda.get_device_properties(i) for i in range(ng)] 41 | s = 'Using CUDA ' 42 | for i in range(0, ng): 43 | if i == 1: 44 | s = ' ' * len(s) 45 | print("%sdevice%g _CudaDeviceProperties(name='%s', total_memory=%dMB)" % 46 | (s, i, x[i].name, x[i].total_memory / c)) 47 | else: 48 | print('Using CPU') 49 | 50 | print('') # skip a line 51 | return torch.device('cuda:0' if cuda else 'cpu') 52 | 53 | 54 | def time_synchronized(): 55 | torch.cuda.synchronize() if torch.cuda.is_available() else None 56 | return time.time() 57 | 58 | 59 | def is_parallel(model): 60 | # is model is parallel with DP or DDP 61 | return type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel) 62 | 63 | 64 | def initialize_weights(model): 65 | for m in model.modules(): 66 | t = type(m) 67 | if t is nn.Conv2d: 68 | pass # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') 69 | elif t is nn.BatchNorm2d: 70 | m.eps = 1e-3 71 | m.momentum = 0.03 72 | elif t in [nn.LeakyReLU, nn.ReLU, nn.ReLU6]: 73 | m.inplace = True 74 | 75 | 76 | def find_modules(model, mclass=nn.Conv2d): 77 | # finds layer indices matching module class 'mclass' 78 | return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)] 79 | 80 | 81 | def sparsity(model): 82 | # Return global model sparsity 83 | a, b = 0., 0. 84 | for p in model.parameters(): 85 | a += p.numel() 86 | b += (p == 0).sum() 87 | return b / a 88 | 89 | 90 | def prune(model, amount=0.3): 91 | # Prune model to requested global sparsity 92 | import torch.nn.utils.prune as prunes 93 | print('Pruning model... ', end='') 94 | for name, m in model.named_modules(): 95 | if isinstance(m, nn.Conv2d): 96 | prune.l1_unstructured(m, name='weight', amount=amount) # prune 97 | prune.remove(m, 'weight') # make permanent 98 | print(' %.3g global sparsity' % sparsity(model)) 99 | 100 | 101 | def fuse_conv_and_bn(conv, bn): 102 | # https://tehnokv.com/posts/fusing-batchnorm-and-conv/ 103 | with torch.no_grad(): 104 | # init 105 | fusedconv = nn.Conv2d(conv.in_channels, 106 | conv.out_channels, 107 | kernel_size=conv.kernel_size, 108 | stride=conv.stride, 109 | padding=conv.padding, 110 | bias=True).to(conv.weight.device) 111 | 112 | # prepare filters 113 | w_conv = conv.weight.clone().view(conv.out_channels, -1) 114 | w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var))) 115 | fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.size())) 116 | 117 | # prepare spatial bias 118 | b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias 119 | b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps)) 120 | fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn) 121 | 122 | return fusedconv 123 | 124 | 125 | def model_info(model, verbose=False): 126 | # Plots a line-by-line description of a PyTorch model 127 | n_p = sum(x.numel() for x in model.parameters()) # number parameters 128 | n_g = sum(x.numel() for x in model.parameters() if x.requires_grad) # number gradients 129 | if verbose: 130 | print('%5s %40s %9s %12s %20s %10s %10s' % ('layer', 'name', 'gradient', 'parameters', 'shape', 'mu', 'sigma')) 131 | for i, (name, p) in enumerate(model.named_parameters()): 132 | name = name.replace('module_list.', '') 133 | print('%5g %40s %9s %12g %20s %10.3g %10.3g' % 134 | (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std())) 135 | 136 | try: # FLOPS 137 | from thop import profile 138 | flops = profile(deepcopy(model), inputs=(torch.zeros(1, 3, 64, 64),), verbose=False)[0] / 1E9 * 2 139 | fs = ', %.1f GFLOPS' % (flops * 100) # 640x640 FLOPS 140 | except: 141 | fs = '' 142 | 143 | print('Model Summary: %g layers, %g parameters, %g gradients%s' % (len(list(model.parameters())), n_p, n_g, fs)) 144 | 145 | 146 | def load_classifier(name='resnet101', n=2): 147 | # Loads a pretrained model reshaped to n-class output 148 | model = models.__dict__[name](pretrained=True) 149 | 150 | # Display model properties 151 | input_size = [3, 224, 224] 152 | input_space = 'RGB' 153 | # input_space = ['RGB'] 154 | input_range = [0, 1] 155 | mean = [0.485, 0.456, 0.406] 156 | std = [0.229, 0.224, 0.225] 157 | for x in [input_size, input_space, input_range, mean, std]: 158 | print(x + ' =', eval(x)) 159 | 160 | # Reshape output to n classes 161 | filters = model.fc.weight.shape[1] 162 | model.fc.bias = nn.Parameter(torch.zeros(n), requires_grad=True) 163 | model.fc.weight = nn.Parameter(torch.zeros(n, filters), requires_grad=True) 164 | model.fc.out_features = n 165 | return model 166 | 167 | 168 | def scale_img(img, ratio=1.0, same_shape=False): # img(16,3,256,416), r=ratio 169 | # scales img(bs,3,y,x) by ratio 170 | if ratio == 1.0: 171 | return img 172 | else: 173 | h, w = img.shape[2:] 174 | s = (int(h * ratio), int(w * ratio)) # new size 175 | img = F.interpolate(img, size=s, mode='bilinear', align_corners=False) # resize 176 | if not same_shape: # pad/crop img 177 | gs = 32 # (pixels) grid size 178 | h, w = [math.ceil(x * ratio / gs) * gs for x in (h, w)] 179 | return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447) # value = imagenet mean 180 | 181 | 182 | def copy_attr(a, b, include=(), exclude=()): 183 | # Copy attributes from b to a, options to only include [...] and to exclude [...] 184 | for k, v in b.__dict__.items(): 185 | if (len(include) and k not in include) or k.startswith('_') or k in exclude: 186 | continue 187 | else: 188 | setattr(a, k, v) 189 | 190 | 191 | class ModelEMA: 192 | """ Model Exponential Moving Average from https://github.com/rwightman/pytorch-image-models 193 | Keep a moving average of everything in the model state_dict (parameters and buffers). 194 | This is intended to allow functionality like 195 | https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage 196 | A smoothed version of the weights is necessary for some training schemes to perform well. 197 | This class is sensitive where it is initialized in the sequence of model init, 198 | GPU assignment and distributed training wrappers. 199 | """ 200 | 201 | def __init__(self, model, decay=0.9999, updates=0): 202 | # Create EMA 203 | self.ema = deepcopy(model.module if is_parallel(model) else model).eval() # FP32 EMA 204 | # if next(model.parameters()).device.type != 'cpu': 205 | # self.ema.half() # FP16 EMA 206 | self.updates = updates # number of EMA updates 207 | self.decay = lambda x: decay * (1 - math.exp(-x / 2000)) # decay exponential ramp (to help early epochs) 208 | for p in self.ema.parameters(): 209 | p.requires_grad_(False) 210 | 211 | def update(self, model): 212 | # Update EMA parameters 213 | with torch.no_grad(): 214 | self.updates += 1 215 | d = self.decay(self.updates) 216 | 217 | msd = model.module.state_dict() if is_parallel(model) else model.state_dict() # model state_dict 218 | for k, v in self.ema.state_dict().items(): 219 | if v.dtype.is_floating_point: 220 | v *= d 221 | v += (1. - d) * msd[k].detach() 222 | 223 | def update_attr(self, model, include=(), exclude=('process_group', 'reducer')): 224 | # Update EMA attributes 225 | copy_attr(self.ema, model, include, exclude) 226 | --------------------------------------------------------------------------------